Conduit agent experiment
A five-stage Go agent pipeline for repository maintenance, with hallucinated-symbol review, compile gates, draft pull-request output, and measured cost and runtime.
Staff Software Engineer · Forward Deployed Engineer · Data & AI · Atlanta
William “MJ” Hill
I’m an Atlanta-based Staff software engineer. I work directly with customers and operators to turn ambiguous requirements into production systems, then stay through deployment. In AI work, I build the evaluation, guardrails, and failure behavior around the model.
Builder and sellerI built the products and closed the sponsorships.
Production ownerStreaming and RSVP systems ran under live event load.
Regulated deliveryThe secure web platform I helped establish serves 48 governments.
13 yrsengineering, incl. 7 under federal contract
$9.9Mfederal contracts supported by my connector work
45public speaking & teaching engagements since 2017
10public curriculum repositories
Choose a path
What do you need?
01Selected work
A few examples, from federal data systems to customer-facing products.
PTSO — ontology-governed multi-agent system
An orchestrator and three specialist agents exchanged data over MCP and A2A but still returned confident, wrong answers. I fixed the coordination problem with an ontology that defines vocabulary, constraints, provenance, staleness, and refusal. Each conclusion carries the rule that produced it, and latency dropped from roughly 20–40 seconds to about 6.
Live demo · Python · MCP · A2A · Google ADK · Neo4j · evidence graph
Space Force Unified Data Library connector
Primary engineer on a connector that streams U.S. and allied sensor data into the Space Force Unified Data Library. Delivered under an Air Force Research Laboratory Phase II award; the work contributed to Meroxa's selection for a $900M IDIQ vendor pool.
Go · Conduit · real-time streaming · ITAR/EAR · Case study
Nuclear Suppliers Group Information Sharing System
Established the web stack foundation for a secure multilateral platform now used by the 48 Participating Governments of the Nuclear Suppliers Group. Named recipient of the National Nuclear Security Administration's Recognition of Excellence in Support of the Nonproliferation Mission.
NNSA / U.S. Department of Energy · multilateral · accredited
S3 log-ingestion Lambda
Built an open-source log shipper around Lambda's 500 MB disk, 2 GB RAM, and 15-minute limits. The streaming design processed a 350 MB compressed file containing several million lines without loading it into memory. I presented the work at AWS re:Invent.
Python · smart_open · asyncio · AWS Lambda · SAR
Higher-education AI under FERPA constraints
Led a live design sprint with Purdue Global stakeholders to adapt the existing STEPs tutoring platform for working adults and other nontraditional learners. We turned the walkthrough into working UI in under an hour and recommended a FERPA-aware path to LMS integration.
Claude Design · LTI 1.3 · stakeholder discovery · FERPA · Case study
NBA Running Net Points
Built an ingestion pipeline and interactive tool for Dr. Justin Jacobs' sports-analytics practice. Standard net-points models estimate possessions from playing time; this one counts the actual possessions in play-by-play data and plots every player's running total across a game.
Live app · Python · Pandas · Flask · Postgres · AWS Lambda · Redis
AI evaluation at Zocdoc
Built the evaluation layer for an automated insurance-matching proof of concept: roughly 100 manually curated ground-truth mappings, confidence-stratified datasets, a Jaccard baseline, and a pipeline for comparing classical, LLM, and hybrid approaches. Separately, I designed an eight-part rubric for repository-level AI coding instructions: identity, conciseness, universal applicability, progressive disclosure, pointers over copies, guardrails, linter boundaries, and a removal test. I used it to ship improvements across two interoperability repositories.
ground truth · baselines · LLM evaluation · Claude Code · C# · Case study
RenderATL sponsorship pipeline
Built a sales agent that finds and scores sponsorship prospects, then queues one lead at a time for review. Policy checks block off-brand language, model calls have batch-level cost caps, and no outreach leaves the system without human approval.
TypeScript · Next.js · Anthropic API · guardrails · Case study
Conference streaming & session RSVP
Built RenderATL's gated streaming product for a 10,000-attendee conference, including ticket entitlements, authentication, playback, and the viewing experience. I also built session RSVP with capacity limits and waitlists. Both ran under live event load.
Authenticated platform · Next.js · Mux · Clerk · Postgres · entitlements
RenderATL Rush
Built a browser game to drive conference awareness and ticket sales. The game used locked leaderboard updates, share-to-unlock rewards, and single-use purchase codes. It generated $7,012.50 in attributed ticket revenue across 11 sales, with a 29.7% code-redemption rate.
Play the game · Phaser 4 · React · Postgres · conversion instrumentation
View the full work index and case studies →
Engineering activity
GitHub activity spans personal and organization-owned repositories. When source code cannot be shared, the work is represented through approved demos, recordings, and case studies instead.
—
contributions in the last year
Mouse over a day, or focus the calendar and use the arrow keys, to see its contribution count.
Less
More
Repository names, commit messages, and organization mappings from private work are never collected or published. Contribution volume is activity, not a measure of impact. View the GitHub profile →
D4BL research agent
A full-stack research agent with authenticated streaming, cited retrieval across vector and structured data, and tracing for model behavior.
PTSO Agent
An ontology-governed multi-agent system that makes provenance, staleness, constraints, and refusal part of the answer contract.
Delivery beyond personal repositories
Production systems and curriculum built with RenderATL and /dev/color are documented through approved case studies and recordings when the underlying repositories are private.
02Talks & workshops
Selected talks and workshops. Recorded sessions are linked below. Browse the complete talk archive → · Open the media kit →
- 2026Commit Your Code Conference — The Put That Sh*t On Agent: Building Multi-Agent AI Systems That Actually Talk to Each Other · Sep 4, 4:30–4:55pm CT · Plano upcoming
- 2026ACM Richard Tapia Conference — Panelist, “Your Brain Trust: How Peer Mentorship Unlocks Harder Technical Problems” · Sep 17, 1:30–2:30pm EDT · Atlanta upcoming
- 2026RenderATL — The Put That Sh*t On Agent: Building Multi-Agent AI Systems That Actually Talk to Each Other — why better models did not fix the problem, and what an ontology contract did (live demo)
- 2026/dev/color MLPD Series — Vibe Coding with Guardrails: Rebuilding MajorKey with Claude Code recorded
- 2026OpenTeams · OSA Community — Can AI Agents Extend the Half-Life of Open Source Software? A Conduit Case Study recorded
- 2026Atlanta Cloud + AI Conference — AI agents for open-source maintenance
- 2025Atlanta Developers' Conference — Beyond APIs: Building Interoperable AI Systems with MCP and A2A Protocols
- 2025/dev/color MLPD Series — Beyond APIs: Building Interoperable AI Systems with MCP and A2A Protocols recorded
- 2025BITCON — Beyond APIs: Building Interoperable AI Systems with MCP and A2A Protocols recorded
- 2025GoWest — Conduit: Real-Time Data Streams Made Simple (and Fast) with Go recorded
- 2025BlackTIDES Data · bidWeek — MCP and A2A workshop recorded
- 2025Mela AI · Secure the Bag & Keep It — AI speaker for an early-career technology event
- 2025Stats On Stats — AI, software engineering, and the future of tech, including NBA x AI: The Ultimate Crossover, a 12-minute RAG demo using NBA data recorded
- 2024BITCON — AI workshop: LangChain, RAG, and AI agents
- 2024The Bean Path — Using AI Agent Workflows to Create Your Personal Research Team — multi-agent orchestration and retrieval-augmented generation recorded
- 2022MagnoliaJS · RenderATL (2021) — You Can Get with
thisor You Can Get With That: Global State Management in React recorded - 2020AWS re:Invent — How the Sausage is Made: using Lambda to send S3 logs to New Relic
- 2020LeadDev — Minimizing risks in code deployment (panel) recorded
- 2019ACM Richard Tapia Conference — Four earlier accepted sessions across 2017, 2019 and 2020, including the peer-reviewed technical panel "Wakanda Name You Got" (10 of 24 accepted)
- 2019Strata Data Conference, O'Reilly / Cloudera — Presented The Will Test during the Better Together Diversity Networking Lunch recorded
- 2017/dev/color in Motion — Moderator, Jeff Dean breakout session
The Will Test
I created this nine-factor framework to help underrepresented engineers judge whether a company can support them. It gives candidates, employees, and leaders a concrete counterpart to Joel Spolsky's Joel Test. I delivered it at five venues between 2019 and 2020.
03Teaching
Ten years designing and teaching technical curriculum. Ten public repositories at github.com/William-Hill.
RenderATL University, "From Vibe Coding to AI Engineering" (2026). I developed the full 34-document curriculum and taught three evening workshops on specs, custom AI workflows, and the guardrails that keep coding agents honest. The series drew 164 registrations, and the first workshop filled to a waitlist.
"Three Tuesday nights. Three workshops. One quiet shift in how I build… I left after each workshop and immediately applied what I learned to refine the platforms I'd been working on."
Oneisha Freeman Rojales · AI & Social Impact Leader
University of the Virgin Islands (2017–2019). I returned as a paid summer instructor for three years, taught on both campuses, and wrote a new curriculum each year. A character-level text generation project from the course later became an accepted technical panel at the ACM Tapia Conference. UVI also invited me to join the faculty as an online adjunct.
Earlier teaching: Senior Adjunct Instructor at Hackbright Academy (2015–2017), paid corporate facilitator for ModelExpand's inclusive recruiting bootcamp, session leader and curriculum designer for the Hidden Genius Project, lead instructor for iSTEAM's 2025 AI and computer vision camp, and volunteer instructor with Black Girls Code and the Kapor Center.
I also founded the Sherry Hill Academic Achievement Scholarship for minority engineering students at Mississippi State. I earned my B.S. and M.S. in computer science there and was inducted into the Bagley College of Engineering Student Hall of Fame.
04Writing
Verified essays and published bylines.
- Models Propose, Contracts Dispose
What a confident wrong answer taught me about ontologies, evidence graphs, staleness, refusal, and where truth should live in a multi-agent system.
- The Will Test — judging how diverse engineering teams are
My original nine-factor framework for evaluating whether an engineering culture can support underrepresented talent.
- Choosing a global state management strategy for your React app
A practical decision framework for matching React state-management tools to an application's actual scope and constraints.
05Contact
Get in touch.
Tell me what you are working on and where you think I can help.