AI cost and reliability, productized
Your AI bill, explained to the token. Then cut. Your answers, kept grounded.
Wyrum finds where your AI spend goes and cuts the waste, and proves your answers stay grounded while the bill falls. My own production system runs at 3.4 cents a question with every receipt published, and the same system catches its own invented facts before they ship. Both are on this page. Go click them.

0
AI tools
~0%
cache hit
0+
KB tables
0+
build cycles
Three services, one niche
An audit that finds the savings, a sprint that ships them, and a retainer that keeps them as models change. Tap any card for the deliverable, the duration, and the guarantee.
1 to 2 weeks on your stack: per answer cost instrumentation down to the token, cache and routing gap analysis, and a grounding review that probes where your model fabricates a citation, misreads a source, or cites something that does not support the claim. Built for legal AI first, it works the same on any agent, RAG app, or support assistant. You leave with a ranked savings plan your team can run with or without me. The same teardown I publish for my own production system, run on yours.
- You get:
- A written teardown with per answer cost waterfalls, the cache and routing gap analysis, a grounding-failure report, and a ranked savings plan.
- Duration:
- 1 to 2 weeks
- Guarantee:
- Full refund if the audit identifies less than 30 percent in verifiable savings.
Proof: My own production system is instrumented in public: 3.4 cents mean per answer, receipts you can check on this site. The method is corpus-independent: the apparatus that catches a fabricated figure catches a fabricated citation, only the database changes.
2 to 4 weeks on your stack: prompt-cache the stable scaffolding (jurisdiction rules, citation formats, your clause library, hot statutes for a legal stack, the equivalent hot context for any other), route cheap models to extraction while frontier models do the hard reasoning, and tighten token budgets and output discipline. Then I install an eval harness, citation-exists, citation-supports, claim-has-source, and hallucination-rate checks, tuned to your domain (holding-accuracy and clause checks for legal, the same checks for any RAG or agent), wired into CI so a deploy that raises your hallucination rate is blocked before it ships.
- You get:
- Shipped optimizations in your codebase, citation-grounding guardrails, and an eval report showing cost and hallucination rate before and after.
- Duration:
- 2 to 4 weeks
- Guarantee:
- Money back if verified savings do not beat the fee within 90 days.
Proof: Techniques proven on a live production agent: about 95 percent cache hit measured, and a published counterfactual showing the same answer at 1.62x the cost without caching, plus a public grounding teardown. The citation-checking apparatus is the same, only the corpus changes.
Continuous monitoring surfaces hallucinated citations, unsupported claims, grounding regressions, drift, and model and price changes to a human ML engineer who reviews and acts. Every month: evals on your real traffic, regression alerts, model and price change reviews, and re evals before any migration. For teams selling into the EU, the same logs become your EU AI Act post-market-monitoring evidence, the Article 11 trail.
- You get:
- A monthly eval report, grounding-regression alerts, model and price reviews, migration recommendations, and a maintained AI Act evidence pack.
- Duration:
- Monthly, ongoing
- Guarantee:
- Tightly scoped monthly deliverables, cancel at term.
Proof: The engineer who built this live production system runs your watch. The Model Watch teardown on this site shows, side by side, where an unwatched model invents facts and where database grounding catches it, the same method, on your case law, your contracts, or whatever corpus your product answers from.
Most AI systems never leave the demo stage. The hard part is everything after: drift, latency, evals, cost, and keeping it reliable. That is the part I have actually shipped.
The proof
Agrotus
Challenge
Build a production AI platform for farming end to end, solo. Grounded agronomy that does not hallucinate, real satellite and vision pipelines, multilingual, secure, and cheap enough to run on one small server.
Solution
A Claude powered assistant with 45 tools over an 80 plus table knowledge base, a custom trained disease model, Sentinel 2 analysis, yield forecasting, and a full operations cockpit. One streaming call per message, dual breakpoint caching, hallucination guards.
Result
Live in production. About 95% cache hit, 3.4 cents mean per answer measured, EN and LT. Designed and shipped solo in about 3 months, across 50 plus build cycles.
Which part proves which niche


The moat: grounding
The same model, with and without grounding. Spot the invented facts.
Cheap is easy to fake. Correct is not. The production system holds every regulatory date, threshold and figure to a citation support check: the answer must resolve to a database record, or it does not ship. No record, no number. Below is the same model answering the same questions with that grounding on, and with it off. Tap the claims you think the ungrounded model invented, then reveal the database record that catches each one.
round 1 of 4
tap the claims you think the ungrounded model invented, then reveal
In Lithuania's nitrate vulnerable zones, what are the autumn and winter date restrictions when spreading nitrogen fertiliser, manure and slurry is banned? Give me the exact prohibited periods.
grounded production answer
measuredBased on Lithuanian NVZ (Nitrate Vulnerable Zone) regulations, last updated 25 April 2026, there is a main closed period that applies to all nitrogen inputs. Mineral nitrogen fertilisers are prohibited from 15 November until 20 March, a span of roughly four months. The same prohibition applies to solid organic manure, which is banned from 15 November until 20 March, and to slurry and liquid manure, which are likewise banned from 15 November until 20 March.
The organic nitrogen cap is a maximum of 170 kg N/ha per year from livestock manure, which is the EU Nitrates Directive limit.
same model, tools disabled, for comparison
Mineral Nitrogen Fertilisers: - Manure and Slurry (organic fertilisers): -
Measured June 2026 on the live production system, Claude Sonnet 4.6, across the eight published showcase questions. Method: same questions, same model, direct API call, tools disabled, no farm data, for comparison. ungrounded answers shown as verbatim excerpts with typography normalized; full captures are committed in the repo. On the farm data question (Q8) the ungrounded model refused outright rather than inventing records; two questions (Q3, Q7) returned empty answers and are excluded.
Why this travels
The grounding method is corpus-independent. It is proven here on an 80 plus table structured-knowledge agent where a wrong number is a real error, and the same apparatus that catches a fabricated figure catches a fabricated case citation. Only the database changes.
Real answers, real cost
What it actually answers, and what each one costs
These are real English questions I put to the live system, with the exact model cost of each answer read straight from the backend. Your domain would be different, the method is the same: grounded answers, with the cost measured, not guessed. When the data was not there, it said so instead of inventing it.
You are dealing with Septoria tritici, caused by Zymoseptoria tritici, at BBCH 39 in central Lithuania. The treatment threshold is reached when there is 10 percent or more infection on the third leaf below the flag leaf, the F minus 3 leaf, at BBCH 32 to 37, or alternatively when prolonged wet and humid periods are forecast, which you already have with three days of rain. Your situation clearly meets the threshold. At BBCH 39, with the flag leaf fully emerged, you are at the critical T2 timing, and the flag leaf must be protected. Three consecutive rainy days create ideal spore dispersal and infection conditions for Zymoseptoria tritici, so spraying is strongly justified.
The reasons to act now line up across several factors. The growth stage is BBCH 39 with the flag leaf exposed, which is the critical window. The wet period of three days of rain means high infection pressure. The threshold is met because the prolonged wet period criterion has been triggered. And the spray timing falls at T2, the flag leaf application, which is the highest yield protecting application of the season. LAMMC trial data consistently shows that T2 fungicide applications at BBCH 37 to 39 deliver an extra 0.5 to 0.8 t/ha in wet Lithuanian seasons.
For products in the Lithuanian registry approved across BBCH 31 to 65, three options stand out. Elatus Era combines benzovindiflupyr and prothioconazole, FRAC groups 7 and 3, at a dose of 0.8 L/ha with a pre-harvest interval of 35 days. Ascra Xpro combines bixafen, fluopyram and prothioconazole, FRAC groups 7, 7 and 3, at a dose of 1.2 to 1.5 L/ha with a pre-harvest interval of 35 days. Prosaro combines prothioconazole and tebuconazole, FRAC groups 3 and 3, at a dose of 0.75 to 1.0 L/ha with a pre-harvest interval of 35 days.
A note from your action plan: Elatus Era is short by 1.9 L and Ascra Xpro is short by 2.5 L, so check stock before scheduling. Prosaro is a solid budget alternative if stock is the constraint. On resistance, avoid using FRAC 3, the DMI or triazole group, alone repeatedly, because Z. tritici has documented DMI sensitivity shifts in the Baltics. The SDHI plus DMI combinations, namely Elatus Era and Ascra Xpro, are preferred for high pressure seasons like this one.
As the next step, before booking the sprayer, verify the Delta T conditions, targeting a Delta T of 2 to 8 degrees Celsius, wind below 15 km/h, and temperature above 8 degrees Celsius.
This breakdown is based on LAMMC and VAAT data for Lithuanian conditions, covering the pollen beetle (Brassicogethes aeneus), its economic thresholds, and treatment options.
The economic threshold for the pollen beetle varies sharply by growth stage. At BBCH 51 to 53, the green bud stage when buds are tight, the threshold is just 1 to 2 beetles per plant. At BBCH 55 to 57, when buds are separating and yellow becomes visible, the threshold rises to 3 to 5 beetles per plant. From BBCH 59 onward, when the first flowers are opening, the threshold is 15 beetles per plant. At BBCH 51 you are in the most sensitive window, where the threshold is only 1 to 2 beetles per plant. The crop cannot compensate for bud feeding at this stage, so the bar for treatment is low. Count beetles on at least 20 to 25 plants across the field before deciding.
Pyrethroid resistance in Brassicogethes aeneus is now widespread across the Baltic region, affecting products in IRAC Group 3A. Do not rely on pyrethroids as your primary choice. Among the recommended IRAC mode of action groups, IRAC Group 4A covers the neonicotinoids such as thiacloprid and acetamiprid, which are effective, but you should check pollinator restrictions and must not spray open flowers. IRAC Groups 1A and 1B cover the organophosphates such as chlorpyrifos-methyl, and you should check the current Lithuanian and EU registration status before use. IRAC Group 6 covers the spinosyns such as spinosad, which are softer on beneficials and suitable for IPM programs.
Pollinator protection is critical. Spray in the evening, after 8 PM, when bees are not foraging, and only before flowers open. Never spray at BBCH 60 or later.
The key action points are as follows. Count now, because if you are seeing beetles at BBCH 51, even 1 to 2 per plant justifies treatment. Avoid pyrethroids in IRAC Group 3A, since resistance is confirmed across Baltic populations. Choose IRAC Group 4A or 6, and rotate groups if you treat more than once per season. For spray timing, use an evening application with no open flowers present and wind below 15 km/h. Finally, check your field, and if you want spray window conditions pulled for a specific field, just name it.
For T2 fungicide rotation against Zymoseptoria tritici, the cause of Septoria leaf blotch, the central reason to rotate away from FRAC 3 is that Z. tritici has well-documented reduced sensitivity to DMI (FRAC 3) triazoles across Europe, including the Baltics. Using FRAC 3 again at T2, the flag leaf timing, would apply the same selection pressure twice in the same season and accelerate resistance shifts in the local population.
The recommended rotation is to FRAC 7, the SDHI group, meaning Succinate Dehydrogenase Inhibitors. This is FRAC group 7, and its mode of action is to inhibit mitochondrial complex II, the succinate dehydrogenase enzyme, which is a completely different site of action from FRAC 3. The key active ingredient here is bixafen. In terms of resistance status, FRAC 7 currently retains good efficacy against Z. tritici in Baltic conditions, with no widespread resistance confirmed.
A suitable T2 product from the Lithuanian registry is Siltra Xpro, which contains bixafen at 60 g/L plus prothioconazole at 200 g/L, combining FRAC 7 and FRAC 3. The dose is 0.75 to 1.0 L/ha, and the BBCH application window is BBCH 30 to 69, with T2 at BBCH 37 to 55 being ideal. The pre-harvest interval is 35 days. It targets Septoria leaf blotch (Zymoseptoria tritici), rusts, powdery mildew, and fusarium ear blight.
One important note is that Siltra Xpro still contains a FRAC 3 partner, prothioconazole. This is standard practice: the FRAC 7 bixafen is the primary anti-resistance component, while the FRAC 3 partner is included at a lower effective load. This is not the same as applying a solo FRAC 3 again.
Summarising the resistance management across timings, T1 at BBCH 32 is already done and used a FRAC 3 triazole applied solo, which is a DMI mode of action. The upcoming T2 at BBCH 37 to 55 should use FRAC 7 (SDHI) plus a FRAC 3 partner, so the SDHI is the primary mode of action and the DMI is secondary.
The key principles applied are these. Use a different primary FRAC group at each timing. The SDHI in FRAC 7 provides strong curative plus protectant activity on the flag leaf, which is the most critical tissue. Mixing modes of action, the 7 plus 3 combination, slows resistance development in both groups. And do not use a solo FRAC 7 either; always mix it with a partner from a different group.
As a final point on timing, the T2 spray at flag leaf, BBCH 37 to 39, is the single most important spray timing for Septoria in winter wheat. Prioritise this window even if T1 was delayed. LAMMC trial data consistently shows the T2 fungicide giving a yield response of 0.5 to 0.8 t/ha in high-disease-pressure years in Lithuania.
Here is a full nitrogen plan for winter wheat targeting 9 t/ha on loam soil, based on LAMMC guidance for Lithuanian and Baltic conditions. The yield target is 9 t/ha, the soil type is loam (medium nitrogen mineralisation), and the total nitrogen recommended is 180 to 200 kg N/ha. Note the NVZ limit in Lithuania, which is 170 kg N/ha as standard and up to 200 kg N/ha with a permit. If you are in a Nitrate Vulnerable Zone (NVZ), check your permit, because 9 t/ha targets typically require a derogation above 170 kg N/ha.
The nitrogen is split across three applications in an R10 structure. The first application, N1, goes on at BBCH 20 to 25 (tillering), in early spring once soil temperature is above 5 degrees Celsius, at a rate of 60 to 80 kg N/ha, to drive tiller activation and early biomass. The second application, N2, goes on at BBCH 31 to 37 (stem extension), roughly 3 to 4 weeks after N1, at 70 to 90 kg N/ha, for stem and spike development. The third application, N3, goes on at BBCH 37 to 45 (flag leaf), at flag leaf emergence, at 40 to 50 kg N/ha, to build grain number and protein content. These three together come to approximately 180 to 200 kg N/ha.
For products, at N1 (BBCH 20 to 25) use YaraBela Axan NS 27-4 (ammonium sulphate nitrate) at 220 to 300 kg/ha, which also supplies sulphur at 15 to 20 kg S/ha, critical on loam soils. At N2 (BBCH 31 to 37) use CAN 27 (Aukso vidukas) or AN 33.5, which are granular and easy to spread. At N3 (BBCH 37 to 45) use KAS-32 or UAN 32 (liquid), which is ideal for late foliar uptake and can be tank-mixed with the fungicide T2.
A few key practical notes follow. Sulphur is essential, because loam soils in the Baltics are frequently sulphur-deficient; target 25 to 35 kg S/ha in total, supplied mainly via N1 with ASN or Axan. The timing of N1 matters most: apply when soil temperature reaches 5 degrees Celsius and no frost is forecast for 48 hours, since early N1 on loam gives the best tiller retention response. Adjust N2 downward if spring is dry, because loam holds moisture but uptake stalls below 30% soil moisture, and consider splitting N2 into two smaller doses.
On the quality side, N3 is the protein dose: if you are growing milling wheat with a protein target above 13%, do not skip N3, as 40 to 50 kg N/ha at flag leaf is the most cost-effective protein booster. For micronutrients, at BBCH 25 to 32 consider foliar manganese at 0.5 to 1.0 L/ha of YaraVita Mantrac on loam soils with pH above 6.5, where manganese availability drops.
Based on Lithuanian NVZ (Nitrate Vulnerable Zone) regulations, last updated 25 April 2026, there is a main closed period that applies to all nitrogen inputs. Mineral nitrogen fertilisers are prohibited from 15 November until 20 March, a span of roughly four months. The same prohibition applies to solid organic manure, which is banned from 15 November until 20 March, and to slurry and liquid manure, which are likewise banned from 15 November until 20 March. Across all of these input types, spreading resumes on 21 March at the earliest.
There is one limited exception, and it applies to grassland only. On grassland with slopes of 5% or less, a restricted application is permitted in two windows. From 15 to 30 November you may apply a maximum of 40 kg N/ha, and from 1 to 20 March you may again apply a maximum of 40 kg N/ha. This exception does not apply to arable land.
Several additional key rules also apply. Organic manure spread on arable land must be incorporated into the soil within 24 hours. A 10 m perennial grass buffer is mandatory where slopes of 2% or more face a watercourse, and no fertiliser inputs are allowed in that buffer. There is a minimum 5 m setback from any water body. The organic nitrogen cap is a maximum of 170 kg N/ha per year from livestock manure, which is the EU Nitrates Directive limit.
On record-keeping, the e-journal (e-zurnalas) has been mandatory since 1 September 2024 for farms of 10 ha or more, and the fine for non-compliance is 750 euros.
In summary, spreading is allowed from 21 March to 14 November, and spreading is banned from 15 November to 20 March. As a practical note, these dates apply across all Lithuanian NVZ zones. If you farm on light or sandy soils or near watercourses, additional local restrictions may apply. Always verify current-year dates with ZUIKVC or NMA, because ministerial orders can adjust them annually.
Under the 2026 Lithuanian CAP rules, the mandatory GAEC conditionality standards are what you must comply with, and non-compliance leads to payment cuts. The standard most relevant to you is GAEC 7 on crop rotation on arable land, which carries a minor penalty level. It requires that at least 35% of your arable land has a different main crop compared with the previous year, and it does not allow the same crop on the same field for two or more consecutive years, while crop mixtures count as a separate crop for the diversification tool. The key 2026 change is that Lithuania relaxed this threshold from 65% down to 35% of arable area, which is a significant easing, but the rule that you cannot grow the same crop two years in a row on the same field still applies.
GAEC 6 on minimum soil cover, which deals with cover crops and winter cover, also carries a minor penalty level. It requires no bare soil from 15 November to 1 March, a window shortened from 15 March in 2026. Acceptable cover includes stubble left unploughed, winter crops, or a sown cover crop, and organic farms are not controlled under GAEC 6. GAEC 5 on erosion risk management, which affects cover crop timing, is likewise a minor penalty. On slopes greater than 12% gradient and at least 20 m in length, row crops such as potatoes and beet are prohibited, and soil protection cover is required from 15 November to 1 March, the same window as GAEC 6, with the new 20 m slope-length criterion added in 2026.
GAEC 8 on landscape features and non-productive areas carries a minor penalty level, and there is good news here: the former 4% non-productive area requirement on arable land has been abolished under an EU regulation change in 2024, applied in Lithuania from 2025 and 2026, so you no longer need to set aside fallow or buffer strips to hit a 4% target. You must still maintain and declare existing landscape elements such as tree groups, ponds, and hedges, and no cutting or clearing is allowed from 1 March to 1 August during the bird nesting season.
Several other GAEC standards are not rotation-specific but carry major penalties. GAEC 1 requires maintaining the permanent grassland ratio within 10% of the 2018 baseline and carries a major penalty. GAEC 2 bans any new drainage in wetlands and peatlands, a ban in force since 2024, and carries a major penalty. GAEC 3 prohibits stubble burning, which the NMA monitors via Copernicus satellite, and carries a major penalty. GAEC 4 requires buffer strips of at least 3 m along rivers and lakes and at least 1 m along drainage ditches shorter than 3 km, and carries a major penalty.
Beyond the mandatory standards, there are voluntary CAP eco-schemes that pay extra on top of basic income support, and these are optional but financially attractive. The Crop Rotation eco-scheme, Augalu kaita, asks for stricter rotation than the GAEC 7 baseline and pays an extra 30.10 euros per hectare. The Cover and Catch Crops eco-scheme, Tarpiniai paseliai, requires you to sow a cover crop after the main crop harvest and maintain it through winter, and pays an extra 139 euros per hectare. The Perennial grass strips eco-scheme, Daugiameciu zoliu juostos, requires establishing grass strips within arable fields and pays an extra 170 euros per hectare. The short-lived flower or pollinator strips eco-scheme requires sowing annual nectar strips and pays an extra 180 euros per hectare.
For best return on investment, the cover crop eco-scheme at 139 euros per hectare is the highest-value rotation-linked option for arable land. Sow a legume or grass mix, for example phacelia, vetch, or clover, after cereals, since it also satisfies the GAEC 6 soil cover requirement.
As a practical checklist for your farm: rotate at least 35% of your arable area by changing the main crop on those fields compared with 2025; make sure no field grows the same crop two years running by checking your field-by-field history; keep no bare soil from 15 November to 1 March by leaving stubble or sowing a cover crop; do not burn stubble, since the NMA uses satellite monitoring; maintain 3 m buffer strips along all water bodies; and consider the cover crop eco-scheme at 139 euros per hectare, sowing after harvest and declaring it in the ISOBUS or NMA portal.
Looking at the spray window for Laukas 1 over the next three days, here is the full picture combining both data sources. On June 10 the temperature runs 14 to 21 degrees Celsius with wind at 3 m/s (11 km/h), 2.4 mm of rainfall, and a Delta T of roughly 7.5 degrees Celsius, which makes the day borderline because the Delta T is high. On June 11 the temperature is 12 to 17 degrees Celsius with wind at 6 m/s (22 km/h) and a heavy 21.0 mm of rainfall, so the status is do not spray. On June 12 the temperature is 10 to 17 degrees Celsius with low wind, only 0.6 mm of rainfall, and a Delta T of 2.2 degrees Celsius in the morning and 4.6 degrees Celsius in the afternoon, making it the best day.
The best window is June 12 from 06:00 to 10:00. Across the four key parameters, the Delta T is 2.2 degrees Celsius against a target of 2 to 8 degrees Celsius, the wind is roughly 10 km/h against a target of below 15 km/h, the temperature is roughly 10 to 13 degrees Celsius against a target of above 8 degrees Celsius, and the rain risk is just 0.6 mm (low) against a target of dry conditions. All four parameters fall within the optimal range, so this is your green window.
As for why the other days are ruled out, on June 10 the Delta T is currently 7.5 degrees Celsius, which is at the upper limit, and afternoon temperatures hit 21 degrees Celsius, so evaporation risk is high and droplet drift is likely. That makes June 10 borderline at best and not recommended for fungicide. On June 11 there is 21 mm of rain forecast along with wind at 22 km/h, so spraying would be wasted and conditions are unsafe. Skip June 11 entirely.
The recommendation is to spray on June 12 between 06:00 and 10:00. Morning conditions are ideal, with low Delta T, calm wind, and cool temperature. The afternoon window from 17:00 to 20:00, with a Delta T of 4.6 degrees Celsius, is acceptable as a backup if you cannot start early.
One caution to keep in mind: Elatus Era is flagged as short by 1.9 L in your action plan, so confirm stock before scheduling the application.
The work log does not include field names directly, but the entries can be matched by area in hectares to the corresponding fields, so here is a full season summary of all logged field operations across the farm for 2026. Across eight fields the operation counts break down as follows. Laukas1, winter wheat at 39.58 ha, had 2 sprayings, 1 fertiliser application, 2 monitoring visits, 1 soil preparation, and no sowing, for a total of 6 operations. Laukas 2, winter wheat at 7.91 ha, had 2 sprayings, 2 fertiliser applications, no monitoring, 1 soil preparation, and no sowing, for a total of 5. Laukas 3, spring barley at 2.50 ha, had 1 spraying, 1 fertiliser application, 1 monitoring visit, 1 soil preparation, and 1 sowing, for a total of 5. Laukas 4, winter wheat at 9.06 ha, had 2 sprayings, 2 fertiliser applications, 1 monitoring visit, no soil preparation, and no sowing, for a total of 5. Laukas 5, winter wheat at 3.52 ha, had 2 sprayings, 2 fertiliser applications, and nothing else, for a total of 4. Laukas 6, winter wheat at 1.89 ha, had 1 spraying and 2 fertiliser applications, for a total of 3. Laukas 9, winter wheat at 8.03 ha, had 1 spraying and nothing else, for a total of 1. Laukas 10, winter wheat at 1.56 ha, had no operations at all, for a total of 0.
The spraying applications by field were as follows. On laukas1, Axial Plus herbicide was applied on 06 April at 0.9 l/ha, and Prosaro 250 EC fungicide on 13 April at 0.8 l/ha. On laukas 2, Prosaro 250 EC was applied on 14 April at 0.8 l/ha and again on 09 May at 1.0 l/ha. On laukas 3, Mustang Forte herbicide was applied on 16 April at 0.5 l/ha. On laukas 4, Prosaro 250 EC was applied on 14 April at 0.8 l/ha and on 09 May at 0.8 l/ha. On laukas 5, Prosaro 250 EC was applied on 15 April at 0.8 l/ha and on 09 May at 1.0 l/ha. On laukas 6, Prosaro 250 EC was applied on 15 April at 0.8 l/ha. On laukas 9, Prosaro 250 EC was applied on 09 May at 1.0 l/ha. On laukas 10, none was recorded.
It is also worth noting that on a 13.19 ha field, which is likely laukas 7 or laukas 8 and not in the main list, Pictor fungicide was applied on 18 April, and Karate Zeon 50 CS insecticide together with Bortrac 150 boron was applied on 09 April.
The fertiliser applications by field were as follows. On laukas1, YaraBela Axan 27N was applied on 25 March at 222 kg/ha and on 10 April at 296 kg/ha. On laukas 2, YaraBela Axan 27N was applied on 26 March at 222 kg/ha and on 11 April at 296 kg/ha. On laukas 3, YaraMila Complex 12-11-18 was applied on 10 March at 350 kg/ha. On laukas 4, YaraBela Axan 27N was applied on 27 March at 222 kg/ha and on 11 April at 296 kg/ha. On laukas 5, YaraBela Axan 27N was applied on 28 March at 222 kg/ha and on 12 April at 296 kg/ha. On laukas 6, YaraBela Axan 27N was applied on 28 March at 222 kg/ha and on 12 April at 296 kg/ha. On laukas 9, none was recorded, and on laukas 10, none was recorded.
The most active field is laukas1 with 6 operations. Laukas1, at 39.58 ha of winter wheat, leads with 6 logged operations, consisting of a herbicide spray, a fungicide spray, two fertiliser applications, and two monitoring visits. Its NDVI of 0.844 reflects this investment well.
Two fields need attention. Laukas 9 had only 1 operation, a fungicide spray, with no fertiliser recorded this season; its NDRE is good at 0.705, but it is worth verifying that nitrogen was applied. Laukas 10 had zero operations logged this season, with no spray, no fertiliser, and no monitoring, so it needs immediate attention to check whether the records are missing or whether the field was genuinely skipped.
see the full audit teardown of all 8 answersEach cost is the model inference cost of that single answer.
Built end to end, not just optimized
The cost work sits on top of a full production system
Cutting an AI bill is one skill. Building the agent, the retrieval, the models and the pipelines underneath it is another. This system has all of it, shipped and serving real traffic, not a notebook.
Production RAG agent
A 45 tool agent over an 80 plus table knowledge base, with retrieval at inference time, one streaming call per message, and grounding guards on every numeric and regulatory answer. Shipped and serving live traffic.
Custom trained ML
A nine class crop disease vision model trained from scratch, not an API call, plus a multi signal photo diagnosis path that fuses the image with weather, growth stage and region. Live in the product.
Geospatial pipeline
Sentinel 2 imagery to NDVI to K means management zones to ISOXML export, the full path from a satellite pass to a file a tractor can run.
The engineer who audits and optimizes your stack is the one who built all of this.
How it is built
Enough specifics to trust the work. Verified against the live codebase.
AI
Manager agent loop with typed tool dispatch over Claude Sonnet, dual breakpoint prompt caching at about 95% hit.
Cost control
Split conformal routing and token budgets bound model spend to a few cents per message.
Security
JWT in httpOnly cookies, CSRF tokens, organization scoped queries, a read only public demo enforced by middleware, rate limiting.
Stack
Next.js, React, TypeScript and Tailwind on the front. FastAPI, SQLAlchemy, Pydantic and Python on the back. PostgreSQL, SQLite and pgvector for data and retrieval. Self managed VPS with systemd and nginx.
Let’s build something.
Tell me what your AI costs to run. I will show you where the money goes and what comes back: an audit that finds the savings, a sprint that ships them, or a monthly watch that keeps them. Fixed scope, guaranteed, and the person who built the system does the work. Pricing on request.
AI Cost & Grounding Audit, Optimization Sprint, or Model Watch. Ask for pricing.
Designed and shipped solo in about 3 months, across 50 plus build cycles.

