What I’m building right now
Your restaurant.
A little more conversational.
I’m the founder and solo engineer behind an AI host that helps guests find answers, right on a restaurant’s website.
Explore hireMaitreWhat the host does for guests now, and the ordering and voice work I’m building in beta development.
An embeddable host on the restaurant’s own website, answering from that restaurant’s information.
The questions guests ask before they visit, answered from the restaurant’s imported menu and details.
Restaurants connect Square from the dashboard. It is the path the beta ordering flow is designed to take.
Orders taken inside the conversation, with prices resolved on the server rather than by the model.
Real-time voice with live transcription. The intended flow goes from a spoken request to a validated order.
Spice Kraft uses the live chat. Voice and ordering are not live there.
Meet the idea behind hireMaitre: turn everyday questions into a more helpful restaurant experience.
A guest asks in their own words and gets an answer from the restaurant’s real menu, then builds the order in the same conversation. Product demonstration; ordering is in beta development.
See the guest sideThe restaurant sees what guests are asking and answers once, in its own words, so the host gets it right the next time.
See the restaurant sideWatch the setup experience, from bringing in restaurant information to preparing the assistant. This is a product demonstration; voice and ordering remain in beta development.
What happens between a guest’s question and a reply the restaurant can stand behind.
Which dishes are nut-free?
Follow the question through the systemWhich dishes are nut-free?
Setup begins with the restaurant’s website. A multi-strategy importer tries a direct fetch, Firecrawl and then a headless browser, finds menu pages and embedded PDFs, and reads scanned menus with GPT vision. Every outbound request passes SSRF and DNS-rebinding guards.
Each restaurant is its own tenant. Postgres row-level security, service-only RPCs and column-scoped update grants keep one restaurant’s menu and settings out of every other conversation.
The host replies from the records retrieved for that restaurant: its menu, allergy notes, hours and parking.
Ordering is in beta development. A conversation can propose an order, but prices resolve on the server and payment webhooks are signature-verified and deduplicated.
Why the host lives in the page instead of behind a contact form, and what that decision costs to run.
Most of the decision happens before anyone walks in: someone opens the site at nine at night with a question about the menu, the hours or the parking. An AI-native product answers it there, in the page, and then hands the guest to the next step, whether that is reserving a table, starting a Square pickup order, calling or getting directions. Adding a chatbot to a contact form would have kept the wait; removing the wait was the whole point.
Every answer costs money, so the product meters the thing it actually sells. One billable guest message is a public widget request that produces a useful answer for the restaurant. Allowances come with the plan, reset once per billing period and do not roll over. Deciding that unit early is what lets a price exist at all.
The whole app, including SSR, APIs, the crawler, cron and the AI pipelines, compiles into one Cloudflare Worker on the global edge network, so a guest message is answered near the guest and there is no single machine to outgrow. The real ceiling is deliberate: durable per-IP and global quotas, plus fail-closed spend budgets that stop the system rather than let it run up a bill. Behind them sits Postgres with row-level security, which is the first piece I would scale. These are the designed limits, not a load-test result.
Guest conversations, including voice transcripts, are deleted ninety days after they are created by a scheduled job. Restaurants are separate tenants at the database level rather than by convention. More than 2,000 automated security and behavior checks run on every commit, covering tenant isolation and cost abuse among others. The product runs as a controlled beta with a limited number of restaurants, which is the honest stage to be at.
Tenant-scoped caching, in-flight request deduplication, parallel database reads and stage-level timing. A diagnosis, not a model swap.
Malformed-field recovery, browser fallbacks, PDF compatibility and rejection of irrelevant page content, measured on a ten-restaurant benchmark on a normal network.
Exact names, aliases, typo matching, variations and ambiguity checks, shared by chat, voice and ordering. Interpretation stays separate from the transaction.
Context truncation, tool-result eviction, token and cache telemetry, and on-demand cart reads keep a long call inside its budget.
Assistant history reconciled against server logs, credential-shaped output redacted, links restricted, and repeat off-topic requests stopped before paid inference.
An in-process PostgREST, Square and OpenAI test backend, Chromium lifecycle tests, property tests and structured timing reports. The count below is a test result, not a production rate.
Webhook cursors, idempotent upserts, deletion handling and duplicate-event checks, with full reconciliation when the structure changes.
Owners review the extracted questions and supply the answers, so the assistant is not learning from its own wrong replies. Embeddings batch in groups of 64.
Each card is one change from the private hireMaitre repository, written from its commit history. The figures are recorded measurements and test results from the time of the work, not fresh benchmarks and not production rates.
Before hireMaitre
Financial systems. Structured data. Work with reach.
Outreach project on GitHub
Asymmetrica Investments AGEvent-driven pipelines at 99.9% service availability.
Confirmed in Francisco Escobar’s recommendation.
I worked across React, TypeScript, Python, PostgreSQL, Redis and AWS on an internal financial platform that connects real-time market data with portfolio information and risk workflows. Its event-driven pipelines process more than 2 million market-data events a day while maintaining 99.9% service availability.
I built a modular order-management and risk-validation engine. Exposure limits and position checks validate each transaction, idempotent processing keeps retries from becoming duplicates, and immutable audit logs record every decision. Docker, GitHub Actions, automated tests, structured logging and alerting support each release.
I built Python workflows that collect company and investment data from websites and REST APIs, then extract, classify and structure it for AI and RAG processing.
The structured profiles feed personalized outreach generated with OpenAI. SendGrid automates delivery, and structured tracking follows investor engagement and campaign activity.
The outreach work delivered more than 1,000 targeted emails. My manager also highlighted my contributions to the commodity exchange platform.
“A sharp, proactive contributor who gets things done.”
Exploring what makes an AI product reliable.
Research. Build. Listen. Improve.
React, Supabase, PostgreSQL, VercelResearch. Build. Listen. Improve.
React, Supabase, PostgreSQL, VercelI contributed to the development and deployment of FinPal, an AI-enabled personal-finance application built with React, Supabase, PostgreSQL and Vercel, supporting implementation, backend integration, testing and release workflows.
I tested voice and video agents through scenario-based testing, documenting issues in conversational consistency, session reliability, knowledge freshness, identity continuity and response quality, then translated the findings into product recommendations.
I researched agent frameworks and integration patterns, including LangGraph, LangChain, real-time speech processing and multimodal interfaces, and co-authored the published report below.
How multimodal AI agents are built and run: LLM orchestration, memory and tool use, voice and video interfaces, speech-to-text and text-to-speech, deployment and scaling.
Read the report on ResearchGateFour explorations in how systems think,
organize, predict and protect.
Type a review. Watch a model decide, and see the words it decided on.
Try the live appRead the code on GitHubHosted on a free instance, so the first request after it sleeps can take about a minute to wake.
Every review arrives as free text with one human label, positive or negative. When the labeled set is big enough, the trainer holds back a stratified slice, scores itself on reviews it did not fit, then refits on everything. The deployed demo ships with ten sample reviews and fits on all of them, so what it shows you is behavior, not a benchmark.
A classifier cannot read. TF-IDF scores each term by how often it appears in one review against how common it is everywhere, so filler words weigh almost nothing. I fit two vectorizers side by side: word unigrams and bigrams, so “was bad” counts as its own feature, and character three to five grams, which keep working through typos and odd spacing.
Plotted, those vectors give the picture the rest of this chapter is built on. Reviews worded alike sit close together and the two labeled groups pull apart. The two axes stand in for a space with one dimension per feature, far wider than any screen; the geometry is the point, not the picture.
You can classify a new review without training anything: find the five nearest labeled reviews and let them vote. k-nearest neighbors is the clearest way to see what similarity means in this space, and it is where my data mining coursework started. It keeps every example instead of learning from them, which is exactly why it is not what I shipped.
Movie Mood Machine does not take a vote. Logistic regression learns one weight per term and adds them up, and that sum is a single boundary through the whole space. A prediction becomes two facts: which side of the line a review falls on, and how far from it.
That distance goes through a logistic curve, which turns it into a probability. The app reports both sides rather than a bare verdict, because a review sitting on the line and one sitting far from it should not look the same to the person reading the result.
The same weights say which terms pushed a prediction, so the app lists them instead of asking for trust. The fitted pipeline is saved with joblib, the container is a slim Python image, and one small JSON endpoint answers the page: post a review, get the label, both probabilities and the signals behind them.
Store detail only where detail exists.
Academic project. Source code available on request.
An 8 × 8 image can be stored as 64 individual pixels. But large areas often contain exactly the same color.
Check whether a region is uniform. If it is mixed, divide it into four quadrants. The animation follows the pattern in my original project diagram.
Mixed quadrants keep splitting. Uniform ones stop. Watch the image partitions and tree nodes build together as the recursion moves deeper.
My Java image engine supports pixel updates, brightness computation and reconstruction, with a custom dynamic circular queue and recursive restructuring.
Next Shift: sales forecasting from real restaurant signals.
Explore Next Shift on GitHubRead-only Square OAuth brings in orders, payments and refunds. The pipeline deduplicates records and aligns them to the restaurant’s business day.
Daily totals become lag values and rolling averages, joined with weather and calendar signals. Pandas transforms the history into model-ready features.
A PyTorch network maps 68 inputs through layers of 64 and 32 units to one sales prediction. Chronological splits and early stopping keep evaluation tied to future performance.
The forecast includes an 80% range and expected order count. It is compared with last week and a seven-day average, with measured results reported even when a simple baseline wins.
The moment AI can take action,
the boundaries matter.
A useful agent may touch customer records, request refunds or stream sensitive information. Every connection needs a clear answer to three questions: who can act, what can leave, and what happens when something fails?
Four focused services for an AI platform engineering assessment.
View all four services on GitHub123 role = ROLE_BY_TOKEN.get(bearer_token(authorization) or "")124 if role is None:125 return rpc_error(rpc.id, UNAUTHENTICATED, "Authentication required")130 return rpc_error(rpc.id, -32602, "Invalid params: tool name is missing or malformed")131 if is_admin_tool(name) and role != "admin":132 log.warning("blocked %s for role=%s", name, role)133 return rpc_error(rpc.id, UNAUTHORIZED_TOOL, "Unauthorized Tool Call")The customer refund MCP service defines a focused tool boundary. Structured requests and validation make the operation explicit instead of treating model-generated text as an instruction to execute.
The security gateway validates JSON-RPC, checks the bearer token’s role and blocks restricted administrative tools. The highlighted excerpt is taken directly from the repository.
The streaming PII guardrail handles sensitive information during output. This brings privacy checks into the response path, where partial chunks and boundaries need deliberate handling.
A token-aware rate limiter and fallback router explore how a service can respond when capacity or a provider becomes unavailable. Together, the four services cover action, access, output and resilience.
The request is forwarded only after its role and tool name pass validation.
Read the source on GitHub123 role = ROLE_BY_TOKEN.get(bearer_token(authorization) or "")124 if role is None:125 return rpc_error(rpc.id, UNAUTHENTICATED, "Authentication required")126127 if rpc.method == "tools/call":128 name = (rpc.params or {}).get("name")129 if not isinstance(name, str) or not TOOL_NAME.fullmatch(name):130 return rpc_error(rpc.id, -32602, "Invalid params: tool name is missing or malformed")131 if is_admin_tool(name) and role != "admin":132 log.warning("blocked %s for role=%s", name, role)133 return rpc_error(rpc.id, UNAUTHORIZED_TOOL, "Unauthorized Tool Call")134135 return await forward(request.app.state.http, rpc, payload, role)B.S. Computer Science
Major GPA 3.7, Dean’s List
Python, TypeScript, React, Java, C, SQL
Cloudflare, AWS, PyTorch, OpenAI