Aleksei Nikolaev

A phone photo of a notebook page can’t be searched or copied — but what’s due is a PDF, not a gallery shot.

Send a photo or scan (web or a Telegram bot); classical CV (OpenCV: page detection, perspective correction, alignment, Sauvola binarisation) cleans the page, qwen2.5vl:7b via ollama classifies the page and optionally transcribes it to Markdown, and EasyOCR ru+en on GPU places word-level coordinates into an invisible layer over a 300 dpi A4 PDF. Every stage and the engine that actually ran are shown live; the PDF is ready before the transcript. Nothing goes to third-party AI services, and copies are deleted 30 minutes after delivery.

• 338 tests
• 4 rounds of independent Codex code audit: 44 findings, 43 fixed, 1 accepted as risk
• on a synthetic holdout (28 files: desk shots, shadows, noise, rotation) — median CER 0.2%, p50 5.6 s/page on a warm model, 1-bit page ≤ 96 KB

The chain says one thing, the internal ledger another — and someone has to find out which is wrong before the shift ends.

Reconciliation is deterministic code: chain events matched against double-entry journals and wallet snapshots, seven discrepancy classes, a balance equation that must close per asset. The model only explains — and every figure, hash and address in its answer is checked against the data before anyone sees it.

• 621 tests
• precision/recall 1.0 across 7 seeds
• ≈ $0.001 per answer
• integer minor units, no floats

A compliance question on EU regulation needs an answer with the exact article — not a confident guess.

The agent walks the structure of the acts (article → paragraph → point, cross-references) instead of similar-looking chunks. Every citation is verified by code against the corpus; if it fails, you get a refusal, not an invention. Each answer shows its cost.

• 2 regulations, 2,554 addressable units
• 455 tests
• citation validator is code, not a prompt

“We have RAG” means nothing until you measure which retrieval actually works on your corpus.

Five strategies on one corpus (lexical, dense with two chunkings, hybrid, hybrid + cross-encoder), 47 labelled questions, Recall/MRR/nDCG, judge-scored faithfulness, cost and latency — reproducible, prompt-versioned. Results shown as they are, including the inconvenient ones.

• Recall@10: 0.61 naive windows → 0.91 hybrid + reranker
• 363 tests
• the first hybrid lost — the cause was found and written up

“What expires in 14 days in the DMZ, and where else does this CA live?” — ops questions that must not be answered approximately.

Go, a single static binary. The agent writes guarded SQL, builds a rotation plan with a deterministic ranker, scans real hosts from a whitelist. Every number, hostname and CA in the answer is checked against the query result; an unrecognised number form blocks the answer. ADRs and C4 in the repo.

• 135 tests; 7 findings from an adversarial audit closed the same day
• ≈ $0.0007 per answer
• from spec to public stand in 5 hours

A legal answer without a verifiable reference to the norm is a liability, not help.

Structural retrieval over articles/parts/points of three federal laws; every reference is code-validated and clickable — it opens the actual text of the norm. Each answer shows its cost in tokens and money.

• 588 tests
• citations only from nodes the agent actually opened
• daily cost dashboard

A developer’s floor chart (an image) and a registry export (a PDF) describe the same apartments differently. One trustworthy list is needed.

Four visible stages: OCR → LLM extraction → grounding check of every record against the source → deterministic matching. A model error shows up at its own stage, not in the final table.

• 321 tests
• local OCR (Qwen-VL on own GPU) — documents never leave the machine
• every record flagged: grounded / not

The analyst is busy and the business needs “which category dropped during the promo” now — without confident nonsense.

The agent reads a semantic layer, writes AST-guarded SQL, asks a clarifying question when the request is ambiguous, refuses outside the domain. Every number in the answer must appear in the query result. Full trace: steps, SQL, tokens, cost.

• 120 tests
• eval set with traps: ambiguous → clarify, off-domain → refuse
• real catalogue + synthetic sales, labelled as such

A shopper asks “which subscription tier fits” — and the answer must come only from the real catalogue, at the real price.

Function calling over a structured index (~300 items, all denominations) instead of free generation: the agent cannot name a product or price that does not exist. Quotas and a token budget built in.

• idea to public production in one day
• 211 tests
• zero invented products, by construction

Invoices and delivery notes arrive as images. They must become verified requisites, not “roughly this”.

OCR → structure → code checks of requisites (tax IDs, totals, dates, cross-sums) → verdict; each stage shows its own output.

• 282 tests
• local vision model, no cloud
• you can see exactly where the model went wrong

An engineer needs the clause, not a paraphrase.

Same principle as the legal stands: agentic retrieval over document structure, deterministic check of every reference, click opens the source clause.

• 313 tests
• a reference the corpus cannot confirm means the answer is withheld

15,000 SKUs; the customer describes a task in words (“what to paint outdoor OSB with”), not an article number.

An agent with search tools over a local catalogue index and FAQ, hand-off to a manager for leads; polite harvesting of public data respecting robots.txt.

• 15,562 products, 738 categories
• 153 tests
• the first stand of the line

A test on different material: a character pipeline (passport, hair, physics, MToon export) and a playable scene — can agents carry this too.

Three.js; assets and animation produced by an agent pipeline under review; for the lab — a playground for generation experiments.

• playable in the browser
• a bridge to the gaming domain in outreach