01
What do you actually build?
Working software that uses AI where it helps and plain code where it doesn't. Customer-facing sites, internal tools that replace spreadsheets, and assistants that answer questions from your real data.
AI systems engineer · New Orleans → Atlanta · Remote
Altrell Gross
I build AI systems that hold up in production — retrieval that is measured, assistants that cannot invent your numbers, and the database, tests and deploys underneath them.
Agent orchestrationRAGEvals & guardrailsNext.js · SupabaseLocal AI
Proof, not adjectives
12Systems shipped
7Live in production
400+Automated tests green
$11.1MClient job value reconciled
What clients ask first
Answered before you have to ask.
02
How do I know the AI won't make things up?
Because I don't let it. On money and records the model only chooses which lookup to run — the numbers come straight from your database. Every figure it reports can be traced to a row you can open yourself.
03
Is my data safe?
Access rules live in the database, not just the screen, so a user cannot read what their role forbids even through the API. Sensitive jobs run entirely on local hardware, where nothing leaves the building.
04
What happens after you hand it over?
You get the repository, the tests, the deploy path, and written runbooks. Everything I ship is verified in a real browser and a real database before I call it done — and I tell you what I did not check.
Capability matrix
The skills hiring managers are paying for in 2026 — and where I have actually used each one.
Agent & tool orchestration
#1 fastest-growing skill on LinkedIn, 2026Designing systems where a language model picks tools and a deterministic program does the work. Includes building the MCP servers those tools live in.
- MCP server design
- Tool/function calling
- Multi-agent review
- Cross-model adjudication
- Semantic routing
- Context engineering
- Claude · Gemini · GPT · Llama
Built a 7-tool MCP server used daily by three different AI clients, plus a router covering 51 capabilities.
RAG & retrieval systems
Top-3 paid AI skill clusterGetting the right context in front of a model — chunking, embeddings, hybrid search, and knowing when retrieval is actually making things worse.
- Hybrid keyword + vector search
- Embedding models (bge)
- SQLite FTS5 · Postgres
- Chunking strategy
- Incremental reindexing
- Grounded citations
- On-premise RAG
90% recall@5 on a 50-question evaluation, against a 44% navigation baseline.
Evaluation & guardrails
The clearest signal of real LLM experienceMeasuring whether an AI feature works instead of assuming it does, and building the deterministic checks that stop a confident model from costing you money.
- Eval design & scoring
- Precision/recall tradeoffs
- Prompt-injection defense
- OWASP · OWASP-LLM audit
- Deterministic output guards
- Human-approval gates
- Audit trails
Caught a transcription bug that would have billed $971 instead of $52.30, and shipped the check that makes it impossible.
Production apps & data
Bridging prototype to productionThe unglamorous half: schemas, migrations, row-level security, payments, deploys, and enough tests that a change on Friday is not a gamble.
- TypeScript · React 19 · Next.js 15/16
- Supabase · Postgres · RLS
- Python · Node · PowerShell
- Stripe payments
- Vercel · Hostinger deploys
- Vitest · pgTAP · Playwright
- CI-grade verification
230 automated checks on a single client app, including one proving anonymous visitors read nothing.
Local AI & cost control
Cost optimization is a 2026 hiring line-itemRunning capable models on hardware you already own, so private data stays put and there is no per-call bill attached to daily work.
- Ollama · local LLMs
- faster-whisper speech-to-text
- Vision models (Qwen-VL)
- FLUX image generation
- VRAM budgeting
- Free-tier model routing
- Token-cost engineering
Three models orchestrated inside 6 GB of VRAM, with a hand-off scheme so none of them spill to CPU.
Interface, motion & brand
What makes the work sellableThe part clients see. Brand systems drawn from a client's own material, 3D and scroll motion that carries meaning, verified on real devices.
- React Three Fiber · WebGL
- GSAP · ScrollTrigger · Lenis
- Design systems & tokens
- Typography
- Accessibility & reduced motion
- Responsive QA at real breakpoints
Six distinct brand systems shipped, each built from the client's own source material rather than a template.
Selected work
Six systems, and the hard part of each.
Live in productionRetrieval engine · MCP server
Own R&D
A search engine for everything I know
The pattern behind any 'ask our documents' feature — and proof I measure retrieval instead of trusting it.
A retrieval engine over 1,668 documents and 7,806 text chunks, combining keyword and semantic search over SQLite FTS5 with bge embeddings. It ships as an MCP server so any AI assistant can query it directly, and reindexes incrementally in under a second.
The hard part
Adding graph expansion — the obvious next feature — measurably made results worse, dropping recall from 90% to 80-88%. I cut it from ranking and kept it for visualization only. Most teams would have shipped it because it demos well.
90%Recall@5, measured
1,668Documents indexed
7,806Chunks embedded
PythonSQLite FTS5bge embeddingsMCPThree.js
Data destroyed at handover
No screenshot exists. The client environment was fully decommissioned at the close of the engagement — hosting and database deleted, and the client’s data with them. The architecture is described here instead.
- Next.js
- Supabase
- Postgres RLS
- Vercel
- Gemini
- Groq Whisper
Delivered · engagement closedJob costing & profit control
Security integration firm
Replaced the spreadsheet running a business
One system of record per job — estimate, purchase orders, bills, shipping, tax and true margin — instead of a workbook only one person could safely touch.
A Next.js and Supabase app that became the financial record for every job. 28 migrations, row-level security on every table, and a tamper-evident append-only audit log. Migrated 106 jobs, 112 catalogue items and 31 invoices, reconciled to the client's own totals to the cent.
The hard part
The built-in assistant never sees a number. It only picks which database tool to run; the figures come back from Postgres under the asking user's own permissions. That makes an invented total structurally impossible, and stops one role reading another's margins.
$11.1MJob value reconciled
22Security findings fixed
28Database migrations
Next.jsSupabasePostgres RLSVercelGeminiGroq Whisper
Live in productionDispatch & invoicing
Property maintenance company
Dispatch for 91 units and five technicians
Work orders, technician routing, time tracking and invoicing for two apartment properties — replacing texts and paper.
Next.js 16 and Supabase, live on Vercel. 91 units and 152 residents imported, three roles, row-level security across 16 tables, and a database trigger that routes every HVAC job to the only trained technician automatically.
The hard part
Technicians pitch their own pay. Management can approve or send an invoice back with a note, but cannot overwrite a technician's number — enforced in the schema, not the UI. Resident names and phone numbers were also purged from git history before the first push.
230Automated checks green
91Units under management
16Tables under row-level security
Next.js 16SupabasePostgres RLSVitestpgTAPVercel
Visit live site
Live in productionWebsite + local AI invoicing
Sarcio Services LLC · founder-operated company
Invoices dictated out loud, priced offline
Speak the job on the drive home, get a priced invoice with a payment link and a sourced pricing sheet — without client details ever leaving the machine.
A public site plus a local-only invoicing tool. Speech-to-text runs on the GPU with faster-whisper, extraction uses a local Qwen model constrained to a JSON schema, and a second vision model reads the job photos. Payment links are minted per invoice through Stripe.
The hard part
An early build misheard prices and produced $971 of materials instead of $52.30. The fix is deterministic, not a better prompt: an amount is accepted only if that exact figure appears in the transcript. Asked to infer a job from photos alone, the model invented work that was never done — so it is now barred from writing photo notes into billable scope.
3Models inside 6 GB VRAM
~2sSpeech to transcript
0Client records sent off-device
Pythonfaster-whisperOllama · Qwen3Qwen-VLStripePillow
Visit live site
ShippedBrand site + on-premise AI proposal
Mabel Solicitors · Lagos, Nigeria
A law firm's brand, drawn from its own cloth
A public presence matching a firm established in 2007, plus an architecture for AI document search that never sends client matters to the cloud.
The landing page was built from the firm's own corporate profile — its motto, practice areas and offices. The visual language borrows from Adire textile, with a scroll-driven gold thread spine and real liquid-glass effects. Background textures were generated locally rather than licensed.
The hard part
For a law firm, cloud AI is a confidentiality problem, not a cost one. The proposed system keeps every document on a firm-owned machine in Lagos running a local model, so privileged material never crosses the network.
13Practice areas mapped
4Offices represented
100%On-premise data path
GSAPLenisSVG filtersFLUX.1-schnellLocal RAG
Working prototypeVision → structured output
Own product
A photo of a wall becomes a build guide
Point a phone at the problem and get a measured, step-by-step plan — including the ways the job typically gets ruined.
A photo plus a short description is turned into a full build guide: derived measurements, materials, sequenced steps and failure modes. It is a working demonstration of turning an unstructured image into structured, checkable output.
The hard part
Measurements inferred from a photo are estimates, and treating them as exact is how someone cuts material wrong. The output labels them as photo-derived and tells you to verify before cutting, rather than presenting a confident number.
10+Dimensions derived per photo
±5%Stated accuracy ceiling
Vision modelsStructured outputPython
Also shipped
Sites, stores and interfaces.
Live in production
BD Services
This site
3D neural interface, service pages, intake forms and analytics, exported static.
Visit
Live in production
Dr. Tina Health Initiative
Documented before and after
A public redesign kept as a factual case study, with no invented outcome claims.
Live in production
CZ MD Aesthetics
Physician-led aesthetics
Editorial typography, rose-gold glass treatment, and a booking path.
Working prototype
Bagneris Law · New Orleans
Noir and gold
A scroll-scrubbed video hero and a heavier editorial register for legal work.
Live in production
Chris Maison
Cart tested end to end
Browser smoke tests covering the cart and the www/apex domain split.
Visit
Live in production
Founder-operated subscription project
Recurring billing, live
A Stripe-backed subscription signup flow with eight live price tiers.
Visit
Working prototype
Own R&D
An agent with a desktop
A standalone transparent desktop interface driving a local agent runtime.
How I work
Five rules that decide what ships.
01
Measure before believing
AI features get an evaluation with a score, not a demo. If a change does not beat the baseline it does not ship, even when it is the interesting one.
02
Deterministic where it counts
Money, records and permissions are handled by code with checks, never by a model's judgement. The model is allowed to choose; it is not allowed to compute.
03
Security in the database
Access rules live in Postgres row-level security, so they hold through the API and not only behind a screen. Audit logs are append-only.
04
Verified in a real browser
Nothing is called done from a passing build alone. Pages get rendered and looked at across real breakpoints before handoff.
05
Honest handover
You get the repo, tests, deploy path and runbooks — plus a plain list of what was not checked and what is known to be rough.
“Evidence before assertions.”
If I say it works, I ran it and looked at the result.
Track record
Where the discipline came from.
AI Systems Engineer & Consultant
BD Services
2026 — Present
- Design and ship AI-assisted systems end to end: retrieval, agents, evaluation, database, interface and deploy.
- Deliver client web platforms, internal tools and grounded assistants for legal, medical, trades and property-management clients.
- Own the whole path — architecture, security review, testing, deployment and written handover.
Maintenance Technician
Brown Maintenance · Stone Mountain, GA
2026 — Present
- Field maintenance across two apartment properties — the domain knowledge behind the dispatch system above.
Videographer Manager
Perfected Innovations
2010 — Present
- Managed filming, photography, editing, archiving and client delivery.
- Built repeatable content pipelines across capture, organization, compression and distribution.
Service Technician
Vivint Smart Homes
2019
- Installed, configured and troubleshot smart-home, camera, security, access and networking systems.
Telecom / Data Technician
Field roles
2012 — 2014
- Installed and terminated Cat5/Cat6, fiber, wireless access points, patch panels and voice/data systems.
Fire Controlman Training
United States Navy
2009 — 2010
- Trained on digital computer equipment, diagnostics, electronics troubleshooting and operational discipline.
Contact
Tell me what is slow, manual, or wrong.
Send the problem in plain language. I will tell you whether AI genuinely helps, what it would cost, and when it is the wrong tool — before you spend anything.