Professional Experience

A local-first desktop cockpit that gives an autonomous coding agent verifiable accountability — a tamper-evident index of everything the agent actually did on the machine.

• Designed an append-only, hash-chained action ledger anchored to observed effects (post-write file hashes, per-run git-HEAD delta) rather than parsed command strings — because a fully-bypassed agent can make the command lie
• A 5-band honesty model with an explicit 'opaque' verdict: an effect it can't derive is flagged, never silently marked safe
• Ships two signed macOS apps from one codebase via build-time tree-shaking; a single git tag triggers CI that builds both
• Root-caused a memory blowup to an 8-second re-parse of a 53 MB transcript, fixed with byte-offset bookmarks (RSS held flat); collapsed the git scanner from 5 calls/repo to 2 and parallelized it — ~40 repos went 3s → 150–300ms

A Project-Intelligence OS: RAG + knowledge graph over enterprise context, answerable in natural language with an immutable audit trail of what the AI did.

• Chose Postgres + pgvector over Neo4j / Pinecone to keep relational and vector data in one store with native joins and $0 infra overhead — logged with the recall threshold that would falsify the choice
• Picked HNSW over IVFFlat for incremental inserts without a training step; made the chunk (not the document) the retrieval unit
• Origin of the hash-chain primitive later re-implemented in Rust in the Cockpit

Turns call recordings into searchable Japanese summaries + a structured analytics warehouse, so purchasing staff stop re-listening to whole calls for context.

• Built an event-driven pipeline (S3 → SQS → SageMaker async → SNS → batch) running self-hosted Whisper Large-v3 + NeMo diarization plus structured extraction — insourcing ML deleted external AI-vendor dependencies, a privacy win for a business handling customer PII
• Cut LLM cost with in-context batching (Pydantic-array output ≈ 80% fewer prompt tokens) and treated the model as untrusted: rejecting transcripts whose timestamp count drifts outside a bounded ratio
• Decomposed a large monolith into least-privilege read/write/feedback microservices; documented the async-202-poll pattern with its real gotchas (SQS visibility buffer, DynamoDB eventual-consistency)

An omni-channel (LINE / WhatsApp / Telegram / Web) multi-agent system that buys back used electronics end-to-end — dynamically discovering which specs matter per product instead of hardcoding question trees.

• Fused LLM flexibility with deterministic domain truth: the model proposes a per-product schema, then a hand-curated hardware-spec table forbids impossible fields
• LangGraph supervisor + specialist nodes; structured output everywhere with retry/backoff
• A real P99 tail-latency investigation (45s = 11.5× P50) drove schema-cache pre-warming and a circuit breaker; HITL gating on payout

A multi-tenant outbound/inbound call-automation platform: auto-dial on CRM events, throttled bulk campaigns, and a Japanese inbound voicebot for product registration.

• Protected the carrier with an EventBridge → SQS → Lambda rate-limiter and a bounded retry state machine (max 5, then FAILED) that deliberately stops harassing customers
• Voicebot pairs Amazon Lex V2 (ja_JP) with Bedrock Claude Haiku for dynamic question generation; hardened with spam / dead-air / duration guards
• Handled real telephony failure modes: dead air, brand misrecognition, PostCode format; incomplete-save on early hangup

An AI security-review harness where only findings backed by machine-verifiable proof block CI — verdicts are assigned by code, not by an LLM.

• The centerpiece is a pure-Python proof gate that recomputes every verdict and discards whatever an LLM set: a SAST static trace alone returns 'unconfirmable' and must escalate to a live PoC
• Chose LangGraph over CrewAI because CrewAI 'trusts the LLM's word' with no place for a non-LLM decision node
• Drove cost to a <50k-tokens/scan target by letting scanners do the heavy lifting; reports signed with in-toto/DSSE

Two cost-bounded inference products: batch background-removal for listing photos, and SEO-optimized Japanese listing copy with cited specs.

• Photo pipeline: a custom-trained BiRefNet matting model behind a Florence-2 VLM pre-filter that decides whether removal is even needed; GPU endpoints autoscale on SQS queue depth and stop when idle
• Reserved concurrency matches an external API's RPM ceiling (backpressure, not crashes)
• Copy pipeline: JSON-schema-constrained Gemini output (Zod → responseSchema → re-validate) with content-hash prompt caching and in-flight dedupe

Redesigned a large used-goods trading platform from monolithic to microservices, optimizing performance and letting teams ship independently.

• Applied NX Monorepo with explicit boundaries between internal services and public APIs — boundaries as structure, not convention
• NestJS microservices following SOLID with stable contracts; per-environment CI/CD on AWS (EC2 + CodePipeline)
• Event-driven notification architecture (SQS & Lambda), with MySQL query and index optimization

Maintenance and feature development for an application connecting dental clinics in the French market.

• Full-stack development with NestJS, Vue.js, and Vuetify
• Improved UI/UX and fixed bugs based on real user feedback
• Developed asynchronous collaboration skills with a European team

Developed a SaaS platform for creating and sharing interactive 3D assembly instructions, replacing manual 2D documents.

• Integrated 3D display on the web to replace 2D CAD documents
• Converted Figma designs to real interfaces and built backend APIs
• Combined Three.js, Strapi, and NestJS in a Docker environment

Built interfaces for a cryptocurrency market learning and news platform.

• Connected APIs to display on-chain data and supported technical documentation
• Optimized user experience across multiple devices
• Developed learning progress tracking modules