Capabilities → evidence

Deployed a grounded RAG chatbot that passed a 33/33 golden-QA gate before ship.

Shipped a second independent instance of the same pipeline, tuned to a different business's content and brand.

Built a public demo where anyone can watch a real traced pipeline run step through its stages, then run the actual dashboard themselves via a one-command Docker bundle.

Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.

Workflows pause mid-stage for human approval and resume from the gate.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the discovery mode clients actually get. The client-engagement validation is the stronger evidence, and the case study ranks them.

Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.

When the model can't place a response, the pipeline degrades how specific the answer is rather than inventing one: a fallback ladder that climbs the theme hierarchy and then stops, recording the response as unmapped instead of fabricating a finding.

Built a browser-based realtime voice agent across OpenAI Realtime/WebRTC and Gemini Live/WebSockets, with streamed audio, interruption handling, transcription and short-lived session credentials.

Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.

Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.

Deployed a grounded RAG chatbot that passed a 33/33 golden-QA gate before ship.

Keyword and vector retrieval fused by RRF, degrading gracefully rather than failing, and running in the live demos today.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Turns a prospect's URL into a live grounded chatbot and a priced proposal, through 7 human-gated stages.

Extracted and sanitized the legislation RAG built for Stoa: section-level bill corpus, hybrid retrieval, citations a reviewer can verify.

Deployed a grounded RAG chatbot that passed a 33/33 golden-QA gate before ship.

Chatbots must pass the full golden suite before and after deploy.

Picked the shipped model against explicit pass/fail gates on quality, cost, and latency, with the report to show the reasoning.

Generates tailored documents that never invent a fact, with keyword-coverage checks and a human checkpoint before anything ships.

Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the discovery mode clients actually get. The client-engagement validation is the stronger evidence, and the case study ranks them.

Requests fall back through an ordered provider chain on failure, retry against schema validation, and every call's cost is logged automatically.

Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.

Built a public demo where anyone can watch a real traced pipeline run step through its stages, then run the actual dashboard themselves via a one-command Docker bundle.

Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.

Requests fall back through an ordered provider chain on failure, retry against schema validation, and every call's cost is logged automatically.

Picked the shipped model against explicit pass/fail gates on quality, cost, and latency, with the report to show the reasoning.

Derived per-job cost from token mechanics; caught a 6-7x pricing error in code constants.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.

Integrated Bedrock via langchain_aws so structured outputs pass schema validation on arrival, and built a throttling ladder tuned to Bedrock's actual per-minute quota instead of a generic backoff curve.

Requests fall back through an ordered provider chain on failure, retry against schema validation, and every call's cost is logged automatically.

Won't generate outreach copy for a claim it can't back up with proof.

Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.

Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.

Clone → docker compose up → logged-in analysis dashboard in ~5 minutes.

Each engagement runs isolated on shared infrastructure, with automatic TLS and secrets generated per engagement rather than reused.

Proves proprietary code is absent from every published image layer.

Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.

Built and shipped a hand-written MCP server (5 tools, 3 resources, 3 prompts) with auth and per-tool rate limits enforced.

Segmented ~800 recorded sales calls on conversational shape and content, then tested the segments against real conversion outcomes.

Found that tuned ensembles beat a linear baseline by nothing, and shipped the linear model with learning curves as the evidence.

When the model can't place a response, the pipeline degrades how specific the answer is rather than inventing one: a fallback ladder that climbs the theme hierarchy and then stops, recording the response as unmapped instead of fabricating a finding.

Generates tailored documents that never invent a fact, with keyword-coverage checks and a human checkpoint before anything ships.

Won't generate outreach copy for a claim it can't back up with proof.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Generates tailored documents that never invent a fact, with keyword-coverage checks and a human checkpoint before anything ships.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the discovery mode clients actually get. The client-engagement validation is the stronger evidence, and the case study ranks them.

Turns a prospect's URL into a live grounded chatbot and a priced proposal, through 7 human-gated stages.

Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.

Integrated Bedrock via langchain_aws so structured outputs pass schema validation on arrival, and built a throttling ladder tuned to Bedrock's actual per-minute quota instead of a generic backoff curve.

Requests fall back through an ordered provider chain on failure, retry against schema validation, and every call's cost is logged automatically.

When the model can't place a response, the pipeline degrades how specific the answer is rather than inventing one: a fallback ladder that climbs the theme hierarchy and then stops, recording the response as unmapped instead of fabricating a finding.

Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.

Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.

Integrated Bedrock via langchain_aws so structured outputs pass schema validation on arrival, and built a throttling ladder tuned to Bedrock's actual per-minute quota instead of a generic backoff curve.

Workflows pause mid-stage for human approval and resume from the gate.

Mapped my orchestration engine to LangGraph primitives, concept for concept, so the equivalence can be checked directly.

Integrated Bedrock via langchain_aws so structured outputs pass schema validation on arrival, and built a throttling ladder tuned to Bedrock's actual per-minute quota instead of a generic backoff curve.

Derived per-job cost from token mechanics; caught a 6-7x pricing error in code constants.

Authored IaC for a production-shaped product; separately operated an upstream AWS stack.

What I ran on AWS, mapped service-for-service to the Azure AI stack.

Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.

Built the n8n automation layer joining enquiry, CRM, calls, follow-up, enrolment records and Student Services handover across a distributed admissions operation.

Built and shipped a hand-written MCP server (5 tools, 3 resources, 3 prompts) with auth and per-tool rate limits enforced.

Won't generate outreach copy for a claim it can't back up with proof.

Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.

Built the n8n automation layer joining enquiry, CRM, calls, follow-up, enrolment records and Student Services handover across a distributed admissions operation.

Built a browser-based realtime voice agent across OpenAI Realtime/WebRTC and Gemini Live/WebSockets, with streamed audio, interruption handling, transcription and short-lived session credentials.

Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.

Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.

Keyword and vector retrieval fused by RRF, degrading gracefully rather than failing, and running in the live demos today.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Extracted and sanitized the legislation RAG built for Stoa: section-level bill corpus, hybrid retrieval, citations a reviewer can verify.

Built a public demo where anyone can watch a real traced pipeline run step through its stages, then run the actual dashboard themselves via a one-command Docker bundle.

Workflows pause mid-stage for human approval and resume from the gate.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the discovery mode clients actually get. The client-engagement validation is the stronger evidence, and the case study ranks them.

When the model can't place a response, the pipeline degrades how specific the answer is rather than inventing one: a fallback ladder that climbs the theme hierarchy and then stops, recording the response as unmapped instead of fabricating a finding.

Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.

Each engagement runs isolated on shared infrastructure, with automatic TLS and secrets generated per engagement rather than reused.

Proves proprietary code is absent from every published image layer.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.

Deployed a grounded RAG chatbot that passed a 33/33 golden-QA gate before ship.

Shipped a second independent instance of the same pipeline, tuned to a different business's content and brand.

Built a public demo where anyone can watch a real traced pipeline run step through its stages, then run the actual dashboard themselves via a one-command Docker bundle.

Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the discovery mode clients actually get. The client-engagement validation is the stronger evidence, and the case study ranks them.

Turns a prospect's URL into a live grounded chatbot and a priced proposal, through 7 human-gated stages.

Built a browser-based realtime voice agent across OpenAI Realtime/WebRTC and Gemini Live/WebSockets, with streamed audio, interruption handling, transcription and short-lived session credentials.

Built a sovereign consultation-analysis platform solo: pipeline, backend, review frontend, infrastructure and docs; piloted on real consultation data as an ongoing paid engagement.

Turns a prospect's URL into a live grounded chatbot and a priced proposal, through 7 human-gated stages.

Authored IaC for a production-shaped product; separately operated an upstream AWS stack.

What I ran on AWS, mapped service-for-service to the Azure AI stack.

Segmented ~800 recorded sales calls on conversational shape and content, then tested the segments against real conversion outcomes.

Found that tuned ensembles beat a linear baseline by nothing, and shipped the linear model with learning curves as the evidence.

Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.