Enterprise DWH & Data Architect · Senior System Analyst · LLM Engineering

Give a marketplace seller a diagnosis of their listing plus concrete actions, rather than generic AI copywriting.

• Mobile-first PWA over a provider-agnostic LLM backend; prompt, response validator, API specification and evals changed as one unit — a drift between them is treated as a defect.
• Evaluation contour beyond assertions: fixtures for stable inputs, deterministic baselines and inputs that must be rejected, plus content-quality guards for language, placeholder text and category coherence.
• Operator surface with budget hard-stops before provider calls, logging sanitation and rollout readiness that only trusts a full verification run.

LLM output became reviewable engineering: regressions are locked by tests, and release readiness is a measured state rather than a feeling.

Turn a handful of property photos and facts into a publishable vertical video with script and subtitles, with no paid API in the MVP path.

• Contracts-first delivery: every interface versioned as schema plus golden fixture plus a fixture that must fail, with a dependency-free validator reused by the layer above it.
• Event-stream contract fixed by sequence fixtures, including terminal replay on reconnect and an out-of-order event that must be rejected.
• Promotion fence for canonical documents: source path, decision id and content hash in the header, edits only through a new version.

A media pipeline where the boundary between components is enforced by tests, so a backend change cannot silently break the renderer.

Put retrieval-augmented answering into back-office operations where a wrong answer has legal cost, so the assistant must be provably grounded.

• Retrieval evaluated by a deterministic offline gate on an isolated corpus: source hit rate, citation precision, groundedness and a fail-closed answer when the corpus does not contain the fact.
• Every external connector ships as adapter plus mock, with live mode behind explicit flags and read access separated from write access.
• Each mutating endpoint has a read-only pre-flight twin that reveals blockers before side effects; dual approval and requester-approver separation for high-risk actions.

An assistant that says "not found" instead of inventing, and a rollout path where enabling real integrations is a deliberate, reversible switch.

Stop rewriting integrations per project: expose one stable interface per domain and hide the provider and the jurisdiction behind it.

• Identity as a single provider interface normalizing six live providers, with the protocol quirks of each documented rather than discovered twice.
• Unified LLM channel with free and paid tiers, model-array fallback handled by the gateway, zero-retention flags and a demo mode that works with no key at all.
• Payment and billing domains split by jurisdiction behind the same contract, so a project chooses a zone instead of a vendor.

Adding a provider or a country stopped touching product code, and every consumer degrades gracefully instead of failing when a key is absent.

Read structured facts off user screenshots and turn them into a defensible price, with no hand-waved coefficients.

• Three recognition engines behind one record contract with automatic fallback, so the engine can change without touching the frontend.
• Cost control and reproducibility: expensive vision calls cached by image hash, low-confidence values routed to review instead of being written.
• Pricing calibrated by non-negative least squares against actual sales rather than intuition, with a golden set guarding every runtime change.

Recognition accuracy and price quality became measurable, and the strict mode means the service never silently invents a value it could not read.

Walk a non-lawyer from a life situation to a correctly filled document package they can actually submit.

• Situation catalogue with questionnaire-driven document generation, evidence hub and case tracking, content shipped as versioned data bundles rather than hard-coded screens.
• Personal-data compliance built in: consent flow, local PIN lock, revoke-and-delete path.
• Payment channel abstracted behind one purchase interface, with an unavailable-channel fallback so an unsupported store degrades instead of crashing.

Content and monetization can change per market without touching the UI layer, and the build stays green when a payment provider is unavailable.

Stop re-explaining project standards to AI assistants: encode each domain's conventions, traps and acceptance checklists once, in a form an agent loads automatically.

• Twenty-nine skills distilled from real repositories and incident registers, grouped by prefix — warehouse, per-client standards, product domains, cross-cutting discipline.
• Single-source discipline: canonical copies, generated HTML catalogue, mirror synchronization and a re-tiling tool so a shared block cannot drift between copies.
• Every skill carries the reason it exists and the acceptance checklist, so its guidance is auditable rather than folklore.

Onboarding a new engagement or agent takes minutes: the standard, the known traps and the definition of done arrive together.

Convert thousands of hours of a media channel's video output into a structured, falsifiable record of what it asserted and whether the assertions held up.

• Built an end-to-end ingest pipeline (yt-dlp with a Whisper fallback) that harvested and cleaned 927 transcripts and auto-tagged them across 42 concept topics in 13 categories.
• Engineered a hybrid regex-plus-LLM extraction layer that mined 1,054 falsifiable claims and 101 self-reversals, then adjudicated a dated sample into evidence-sourced TRUE / PARTIAL / FALSE verdicts.
• Added an ML calibration and trend layer (embeddings + scikit-learn) that tuned the ranking model to Spearman rho 0.70 and shipped a standalone zero-backend HTML dashboard.

Compresses an unstructured firehose of video content into an auditable, evidence-linked accountability record a non-technical reader can explore offline.

Turn 26 years of raw FX, macro and geopolitical data into an explainable, statistically honest view of what actually moves each currency pair.

• Built a multi-source ingestion pipeline: 26 years of M1 price history for 66 FX and index/commodity instruments (resampled M1 → MN) plus macro series from FRED, ECB, US Treasury and Google Trends.
• Engineered a cross-market correlation engine (14×14 pair matrix and a 78-instrument daily matrix with PCA/factor decomposition, event-study and Granger tests) to rank the real drivers of each pair across 567 geopolitical and ~89K macro-calendar events.
• Wrote a narrative/insight engine that auto-generates plain-language explanations per pair (regime, sentiment, technicals, macro) and served charts, analytics and insights through Flask dashboards.

A reproducible analytics platform that isolates genuine cross-market drivers from noise via factor control, multiple-comparison correction and selection-bias checks.

Make search over a project's accumulated prose — Markdown analyses, SQL, journals, code — deterministic and citable rather than a hallucinated paraphrase.

• Layered retrieval as a ladder — relational metadata for structural questions, then chunk-plus-LIKE search over an addressable chunk store (file / section / line), and a vector layer added only on top when a real chatbot needs it.
• Chunked by semantic boundaries (Markdown by heading, SQL by object, journals per record, code by function); every chunk is self-contained and the reload is idempotent, with per-table assert counts against the sources.
• Hardened the load path against real traps: a control-byte / NUL sanitizer before escaping, client-side case-folding for Cyrillic search, and a "known-garbage returns zero rows" acceptance smoke.

Answers cite file, section and line instead of being recalled; the vector layer stays a recall stage over a deterministic ground truth, never a replacement for it.