Projects

Implementation, and what it saved.

Each study is a system taken from an experiment to something a user or a stakeholder had to trust, with the architecture choice and the cost, time, or review load it moved.

01Live sports broadcast

<15 ms

Broadcast-grade edge vision under 15 ms

Context

International sports tournaments on a single camera feed—millions of live global viewers and a broadcast encode that cannot wait on a distant GPU.

Problem

Single non-stereo camera pose tracking with a hard real-time ceiling under 20 ms and zero perceptual broadcast delay. A stereo rig was not available, and a cloud round trip would have shown up on air.

Camera

One non-stereo broadcast camera. No rig calibration and no sync hardware on the critical path.

Implementation

PyTorch pose and event models were exported to a fixed ONNX graph and compiled with TensorRT for the Jetson Xavier next to the camera. FastAPI only forwarded frames that survived jitter correction and dropout filtering, so the encode path never waited on a training runtime or a cloud GPU.

Impact

End-to-end latency stayed under 15 ms at 200 FPS for live international broadcasts. That kept the stats on the same feed as the picture, which is the difference between a product a rights holder will air and a demo that misses the point.

End-to-end latency: <15 ms
TensorRT throughput: 200 FPS
Live viewers: Global

02FinOps · B2B SaaS

90%+ cut

Multi-agent orchestration and hybrid RAG

Context

Recent FinOps and document platforms I architected—autonomous agents over invoices and financial tables, with GDPR, PSD2, and AML/KYC constraints. A wrong cell is a wrong decision; an open agent loop is a surprise invoice.

Problem

LLM hallucinations, token-cost explosions, and unbounded agent loops over financial tables that do not arrive as clean JSON.

Intake

Unstructured financial tables and documents. A naive chunker invents structure the source never had.

Implementation

Hybrid BM25 plus dense retrieval ran before any generation call. Agent loops in LangGraph-style runners had a hard step cap. LiteLLM sat in front of the providers, and Prometheus recorded cost per query so a routing change showed up as spend the same day.

Impact

Manual invoice and payment handling dropped by more than 90%. Hallucinated answers dropped about 30%, and the gateway stayed above 99% uptime. The cost saving is the review work that no longer needs a person, and the spend that no longer shows up as a surprise on the GPU bill.

Less manual ops: 90%+
Lower hallucination: 30%
Gateway uptime: 99%+

03Supply chain · Logistics

+25%

Document intelligence and automated redaction

Context

Supply-chain and warehousing automation plus multimodal RAG for large enterprise document programs. Invoices and bills of lading arrive in whatever layout the sender used, and PII cannot ride into downstream storage.

Problem

Parsing noisy, multi-layout invoices and bills of lading while scrubbing PII at scale, without parking every document in a human review queue.

Capture

Multi-layout invoices and bills of lading, including scans that defeat template OCR.

Implementation

LayoutLMv3 read the page image and the text together, so a new invoice layout did not need a new template. PII was removed from the pixels with inpainting. Only low-confidence fields were queued for a person. Training and serving ran on SageMaker and ECS, not a notebook.

Impact

Field accuracy rose about 25%, and the human review queue shrank about 70%. That is reviewer time and the cost of a document sitting in a queue, plus PII that never reaches downstream storage.

Field accuracy: +25%
Less human review: 70%
Scrubbed at scale: PII

04Cited answers

Eval gate

Answers that cannot outrun their source

Context

Research and internal-search products where a fluent answer is the product, and a wrong citation is the moment a user stops paying. The hot pattern in 2026 is the same one: hybrid retrieval, a model that writes, and a team that has to defend the answer to a stakeholder.

Problem

The demo answers look specific. On a held-out set of long documents, the model cites a nearby paragraph and states a number that is not in it. Shipping that is a trust problem and a support cost, not a prompt tweak.

Corpus

Long filings, notes, and docs. The experiment starts by holding out questions whose answer is easy to fake from a nearby paragraph.

Implementation

BM25 and dense scores stayed separate. The generator could emit a sentence only when it still pointed at a retrieved span. A held-out question set, including nearby-paragraph traps, was replayed after every mix change. Token spend was traced per answer.

Impact

A fluent wrong citation never reached the user. That protects the subscription: one undefended number is how a research tool gets cancelled. Cost per answer stayed visible, so a more expensive mix had to earn its place on the held-out set.

Required on every claim: Span
Eval set after each mix change: Replay
Visible per answer: Cost

05Production agents

Bounded

An agent that stops before it spends the runway

Context

Support and ops agents are the 2026 default pitch: a model that uses tools and closes the ticket. The failure is an open loop that keeps calling tools, writes a confident update, and blows the GPU bill before anyone notices.

Problem

The POC books a demo by never stopping. In front of a real queue it retries, writes back a status the user did not confirm, and the weekly token bill no longer matches what the founder told the board.

Scope

Ideation starts with the one job the agent is allowed to finish, and the jobs that stay with a person.

Implementation

The agent had one allowed job, a step cap, and a tool allow-list. Write-back required the tool result still attached. Over-budget runs stopped and asked a person instead of buying another call.

Impact

The open demo loop could not become the production path. Token and GPU spend stopped being an unbounded retry, which is the burn line founders actually feel, and a wrong status no longer closed a ticket the user had not confirmed.

Steps and tools bounded: Cap
Write-back stays sourced: Cite
Over-budget runs ask a person: Stop

06Voice agents

Turn-taking

A voice agent that does not talk over the handoff

Context

Voice agents are a 2026 default for support and field work. The demo sounds fluent. The production failure is a turn that writes the wrong job state, then keeps talking.

Problem

Latency, barge-in, and a CRM write are one path. If the model commits a field before the caller has confirmed it, the user and the operator are looking at two different jobs.

Audio

Streaming audio in, partial text out. The experiment is where a partial transcript is still too early to act on.

Implementation

Streaming audio produced a partial transcript, and tool calls waited for a committed turn. The field was read back to the caller before the CRM write. Speech and LLM tokens were capped per call.

Impact

A half-heard answer no longer became the job of record. That avoids a second site visit or a second call, which is the real cost, and a long retry could not run the per-call budget past what the workflow was priced to afford.

Tool calls wait for a committed turn: Turn
Fields confirmed before write: Read-back
Spend capped per call: Budget

07Coding agents

Patch gate

A coding agent that cannot merge its own guess

Context

Coding agents are one of the hottest 2026 products. The POC writes a plausible diff. The MVP has to live in a repo where a wrong patch costs a release.

Problem

The agent edits files it was not asked to touch, the tests it runs are the ones it just rewrote, and a reviewer cannot see which experiment justified the change.

Task

The allowed paths and the definition of done are fixed before the first edit. That is the ideation cut.

Implementation

Allowed paths and a definition of done were fixed before the first edit. The judging tests lived outside the agent's write set. Each patch cited the failing check it answered. Merge stayed with a person, and the agent never held a production credential.

Impact

A plausible diff could not ship itself. Review time went to the cited check instead of a drive-by rewrite, and a bad patch stopped before it became a release incident or a rollback.

Edits stay inside the task: Scope
Tests the agent cannot rewrite: Held-out
Merge stays with a person: Human

08Video understanding

On device

Long video that does not require a cloud GPU to mean something

Context

Video understanding is moving into support, training, and ops tools. The demo uploads a clip to a large model. The product has to run where the footage already is, on a budget the user will keep paying.

Problem

A cloud pass is accurate and too expensive to run on every hour of footage. A small on-device model is cheap and misses the event. The team cannot tell a stakeholder which frames were actually seen.

Sample

Frames are sampled on a schedule. The experiment is which rate still catches the event without watching every frame.

Implementation

A small on-device detector proposed spans. Only those spans went to the heavier model. A label had to point at frames. Low-confidence spans were the only ones a person reviewed.

Impact

The expensive model no longer watched every hour of footage. GPU spend tracked the events, not the raw video, and an operator queue stayed the uncertain tail instead of a full rewatch.

Heavy model sees proposals only: Spans
Labels point at frames: Cite
People review the uncertain part: Tail

09Domain models

Held-out

A fine-tune that has to beat the base model on the user's data

Context

Teams fine-tune because a general model is close but wrong on their nouns. The 2026 failure mode is a leaderboard win that is worse on the tickets, scans, or notes the business actually has.

Problem

The POC looks better on the last twenty examples the founder cares about. On a held-out week of production data it regresses, and the GPU bill for the training run is already spent.

Slice

The data is split before anyone looks at a loss curve. The held-out week is not in the training set.

Implementation

A production week was frozen as the eval slice before training. The untuned model was the baseline. Training had a stop rule. The ship decision compared user-facing errors and serving cost, not average loss.

Impact

A fine-tune that only looked good on the founder's last twenty examples did not get a serving bill. The training run stayed one experiment, and a heavier model had to beat the base model on the locked slice to justify its GPU cost.

Eval locked before training: Slice
Must beat the untuned model: Base
Serving cost is part of the pick: Bill

10Support deflection

Containment

A support agent that knows when to stop

Context

Support agents are sold on deflection and a lower cost per ticket. The user feels it when the bot closes a ticket they still needed a person for. The business feels it when containment goes up and repeat contacts go up with it.

Problem

The POC answers from a help center and marks the ticket done. In production the answer is adjacent, the customer writes back, and the team cannot tell whether the bot saved money or created a second ticket.

Intent

Only a named set of intents is in scope. Everything else is a person, decided before the model is wired up.

Implementation

Only a named set of intents was in scope. An answer had to cite a help-center article. Containment counted only if the customer did not reopen. Escalations carried the cited article to the person.

Impact

A closed ticket that came back was counted as a failure, so deflection could not hide a second contact. That is the support cost the bot was supposed to remove, and the handoff arrived with the article instead of a blank ticket.

Only named intents are automated: Scope
Containment counts a ticket that stays closed: Reopen
Handoffs keep the cited article: Context

Implementation, and what it saved.

Broadcast-grade edge vision under 15 ms

Context

Problem

Implementation

Impact

Multi-agent orchestration and hybrid RAG

Context

Problem

Implementation

Impact

Document intelligence and automated redaction

Context

Problem

Implementation

Impact

Answers that cannot outrun their source

Context

Problem

Implementation

Impact

An agent that stops before it spends the runway

Context

Problem

Implementation

Impact

A voice agent that does not talk over the handoff

Context

Problem

Implementation

Impact

A coding agent that cannot merge its own guess

Context

Problem

Implementation

Impact

Long video that does not require a cloud GPU to mean something

Context

Problem

Implementation

Impact

A fine-tune that has to beat the base model on the user's data

Context

Problem

Implementation

Impact

A support agent that knows when to stop

Context

Problem

Implementation

Impact

Book an architecture review.