mingma.
Open to internships · Summer 2027

Mingma Lama

Mingma Lama — AI engineer and full-stack developer. CS @ Morgan State and AI Engineering Intern at the Applied AI Lab, building grounded AI assistants, multi-agent systems, and a CBT companion robot.

View ProjectsGet in touch

Recent Win · April 2026

Just won a hackathon.

Mingma and the Kairos team accepting the MorganHacks 2026 Second Place check
Mingma holding the $3,000 Second Place check from MorganHacks 2026
Screenshot of the Kairos knowledge galaxy interface
Google Cloud sponsor prize awarded for the MLH Best Use of Gemini API
MorganHacks 2026 · April 12, 2026

Kairos

A metacognitive, persistent knowledge OS

Runner-Up · $3,000MLH Best Use of Gemini API · 1st

Kairos indexes your files, notes, and code into a living knowledge graph powered by Gemini Embedding 2 — then renders it as an interactive galaxy you can orbit. Ask a question and 20+ metacognitive agents fan out across the cosmos, cycle through Recall → Analysis → Insight, and converge into a single answer.

See the project

About

Hello — I'm Mingma.

Mingma Lama portrait

I'm a Computer Science student at Morgan State (graduating May 2027) and an AI Engineering Intern at the university's Applied AI Lab, where I build production AI systems that real students and staff use. My focus is grounding: assistants that answer only from real sources, cite them, and get checked in code rather than trusted blindly.

At the lab I build ORA Navigator, a RAG assistant and NSF proposal reviewer for the Office of Research Administration (ora.inavigator.ai), and before that CS Navigator, an academic advisor for CS students (cs.inavigator.ai). I'm also a research assistant on a NAO humanoid robot that guides students through voice-driven CBT exercises using Claude agents.

On the side I build at hackathons: Kairos, a metacognitive knowledge OS, took Runner-Up ($3,000) + MLH Best Use of Gemini API at MorganHacks 2026, and EmerFlow, a hospital agent swarm for mass-casualty surges, came out of HopHacks 2026. Comfortable across the stack — Google ADK, Gemini, Claude, Vertex AI, FastAPI, Next.js, Postgres + pgvector, Redis, Cloud Run, Three.js.

  • Morgan State University

    B.S. in Computer Science · GPA 3.74 / 4.00

    Graduating May 2027

  • Honors

    Dean's List (CCBC)

    Tau Sigma National Honor Society

  • Baltimore, Maryland

    Open to internships + remote collaboration

Outside the editor

I run.

Outside the editor I run and lift. Just finished my first half marathon, training for a full marathon this October, and an Ironman is on the long list. Endurance work is how I clear my head between sprints.

Mingma at the Naptown Half Marathon finish line

Naptown Half Marathon

Dec 2025 · 13.1 mi

Mingma holding the Baltimore Running Festival 5K crab medal

Baltimore Running Festival 5K

Camden Yards

Baltimore 25th anniversary crab medal in front of the Inner Harbor

25th anniversary medal

Inner Harbor

✓ Half marathonNext: Marathon · Oct 2026Long-term: Ironman

Experience

Where I've worked

  1. AI Engineering Intern

    May 2026 — Present

    ORA Navigator

    Applied AI Lab, Morgan State University · Baltimore, MD

    ora.inavigator.aiGitHub
    • Engineered a production RAG assistant for Morgan State's Office of Research Administration on Google ADK + Gemini with Vertex AI Search over a 751-document knowledge base, plus a three-tier answer cache (in-memory, Redis, semantic embeddings) that cuts repeat Gemini and Cloud SQL calls
    • Designed a three-layer grounding pipeline: a TF-IDF prefetch used only as fact-check context (recall@5 90.4% → 91.7%), Vertex AI Search answering ORA questions only from the knowledge base, and a post-answer check that backfills citations on uncited answers and flags weakly supported ones
    • Designed a proposal reviewer that turns a solicitation into verbatim-quoted requirements and checks a PI's whole package against them and core NSF PAPPG rules — every "met" finding must quote the draft, and the score is computed in code, not by the model
    • Developed Check a Section for single-section PDFs: rule checks, per-rule model review, a three-reader consensus vote, and a grounded proofreader, taking scores from varying per upload to 10/10 identical
    • Scaled reviews to 56-page Research.gov PDFs by splitting sections from the PDF's internal structure, cutting a 300s+ timeout to 58s
    • Built an NSF Form 1030 budget engine (23 PAPPG-cited rules, Excel/PDF export); deployed three services (React 19 PWA, FastAPI, ADK agent) on Cloud Run via Cloud Build CI/CD with Cloud SQL, Redis, Secret Manager, and a ~2,000-test pytest suite
    Google ADKGeminiVertex AI SearchFastAPIReact 19Cloud RunCloud SQLRedispytest

  2. Research Assistant

    Aug 2026 — Present

    NAO CBT Companion Robot

    Applied AI Lab, Morgan State University · Baltimore, MD

    GitHub
    • Programmed a NAO humanoid robot to deliver voice-driven CBT and grounding exercises to students; designed a multi-agent architecture (router, chat, CBT coach, grounding coach, campus RAG) on Anthropic Claude, tiering Haiku for ~1s replies, Sonnet for therapy depth, and Opus for crisis triage
    • Developed the real-time voice pipeline (WebSocket PCM streaming, Silero VAD, Deepgram STT, ElevenLabs TTS) and a pre-dispatch safety gate that intercepts self-harm language before any agent sees it
    • Built speaker-aware turn-taking — the robot turns toward whoever speaks (sound localization + face tracking) and keeps a separate history per person; routed campus CS questions to the CS Navigator RAG API, cutting answers from 10–15s to 4.5–7s
    • Deployed the FastAPI server as an always-on systemd service on a Raspberry Pi with git auto-deploy (~50s push to live) and a live status dashboard (Vercel + Upstash Redis); fixed a cross-user history leak with idle-bounded session scoping and a 14-case pytest suite
    Claude APIFastAPIWebSocketsDeepgramElevenLabsRaspberry PiRedis

  3. AI Engineering Intern

    Jan 2026 — May 2026

    CS Navigator

    Applied AI Lab, Morgan State University · Baltimore, MD

    cs.inavigator.aiGitHub
    • Engineered a production RAG assistant on Google ADK + Gemini 2.5 Flash that answers CS students' questions on courses, prerequisites, degree requirements, faculty, and policies, personalized from their DegreeWorks audit; three modes (CS Navigator, General with Google Search, Coding Tutor) hand off via one-click suggestions
    • Designed a three-layer grounding pipeline: a ~3 ms in-memory TF-IDF prefetch over 70 docs, Vertex AI Search retrieval that answers Morgan questions only from the knowledge base, and a grounding gate (2+ distinct sources or ≥50% claim coverage) plus checks for invented faculty names
    • Developed a Coding Tutor with a sandboxed Python/JavaScript runner (AST blocklists, CPU/memory rlimits, 4s timeout), execution traces, progressive hints, placement quizzes, and mock-interview grading
    • Implemented five-layer conversation memory (rolling summaries, embedding recall on text-embedding-004 + Cloud SQL) with FERPA controls, a three-tier answer cache, and ElevenLabs voice input; cut cold start ~30s → ~10s and answer latency ~15s → ~5s
    • Shipped a Next-Semester Planner (DegreeWorks → prerequisite graph → conflict-free schedules) fed by a 5-min live Banner seat scraper, plus Canvas deadline reminders, a grade-target calculator, and scholarship search
    • Deployed on Cloud Run with Cloud Build CI/CD, Cloud SQL, Secret Manager, and Redis; automated nine Cloud Scheduler jobs for seat refresh, memory consolidation, and reminders
    Google ADKGemini 2.5 FlashVertex AI SearchCloud RunCloud SQLRedisElevenLabs

  4. Web Developer

    Jun 2025 — Aug 2025

    Venetian Italian Eatery & Valentino's Restaurant Group · Edgewood & Baltimore, MD

    • Redesigned and maintained mobile-responsive websites for multiple restaurant locations, improving page load performance, managing menu, promotion, and media content updates, and driving more direct online orders
    • Integrated web updates into Toast POS workflows with ownership, configured and troubleshot POS systems to streamline order and payment processing for staff, and supported web infrastructure to minimize peak-hour downtime
    HTML/CSSJavaScriptToast POS

Selected work

Projects

knowledge galaxy
Project · 2026Featured

Kairos

A persistent, metacognitive knowledge OS

Kairos indexes your files, notes, and code into a living knowledge graph powered by Gemini Embedding 2 — then renders it as an interactive galaxy you can orbit. Ask a question and 20+ reasoning agents fan out across communities, cycle through Recall → Analysis → Insight, and converge into a single answer that could only come from a system that genuinely understands you.

Runner-Up · $3,000· MorganHacks 2026Best Use of Gemini API · 1st· MLH
  • Indexed concepts with Gemini Embedding 2 + tree-sitter AST parsing; ran Leiden community detection on 1,000+ nodes to surface knowledge clusters as planets
  • Orchestrated 20+ metacognitive agents (Nova, Sage, Void, …) via Gemini function calling — each fans across the galaxy in parallel and runs a 3-phase Recall → Analysis → Insight loop
  • Hybrid retrieval (pgvector + tsvector + RRF fusion) grounds every answer; 4-layer persistent memory consolidates nightly so the system evolves across sessions
Next.jsFastAPIThree.jsGemini Embedding 2Gemini 2.5 Flashpgvectortree-sitter
GitHub
ER capacity · live board
Project · 2026

EmerFlow

A hospital agent swarm for mass-casualty surges

A hospital command board where 10 department AI agents (ER, ICU, step-down, OR, staffing, imaging, X-ray, lab, blood bank, EMS) with distinct personas negotiate patient placement in the open, and a coordinator agent writes one plan per cycle. Agents pick units; code owns bed arithmetic and every hard rule.

  • Engineered the rule layer in code: a single move pipeline (validator → apply), a fast lane that places critical patients instantly, a 5-level escalation ladder with hysteresis, and human approval for big actions like ambulance diversion
  • Per-call timeouts and a circuit breaker fall back to rule-based answers so the board never stops, even offline
  • FastAPI + SSE engine with a simulated clock, per-agent memory (5-hour fade) rendered as a 3D force-directed graph in Three.js, a Baltimore ER-capacity EMS map in Next.js 16, and Cloud Run deploys via GitHub Actions backed by a 95-test pytest suite
FastAPIGemini 3.6 FlashVertex AINext.js 16Three.jsCloud Run
GitHub
CDC PLACES 2025
Project · 2026

Mental Distress Prediction Model

Depression + financial stress → mental distress

A state-level predictive model on CDC PLACES 2025 data (50 U.S. states + D.C., n=52) testing whether depression and financial stress jointly predict frequent mental distress. Selected via competitive abstract review for an in-person poster at Morgan State's 31st Annual Undergraduate & Graduate Research Symposium (April 13, 2026).

  • Reshaped CDC PLACES long-format survey data with Pandas pivot tables and imputed missing values via column means across all 50 states + D.C.
  • Engineered a composite "financial threat" feature averaging four CDC hardship indicators — housing insecurity, utility shut-off threat, food insecurity, and food-stamp receipt
  • Trained Scikit-learn LinearRegression on an 80/20 hold-out split and benchmarked against RandomForestRegressor, reporting R² and RMSE on the test set; final fit: Mental Distress = 6.163 + 0.313 × Depression + 0.277 × Financial Threat
PythonPandasScikit-learnMatplotlibNumPy
GitHub 
def generate_blog(transcript: str, voice_samples: list[str]) -> str:
    system = (
        "Mimic the writer's sentence rhythm, vocabulary, "
        "and signature phrases. Cite 2–4 timestamps as "
        "clickable youtube.com deep-links. Pull 1–2 quotes."
    )
    return gemini.generate_content(
        model="gemini-2.5-flash",
        system_instruction=system,
Project · 2026

Lamarca

YouTube videos → blog drafts in your voice

A full-stack Django + PostgreSQL app that turns any YouTube link into a polished blog draft in ~30 seconds. Gemini 2.5 Flash handles generation, Supadata feeds transcripts, and a voice-cloning prompt mimics each user's sentence rhythm and signature phrases — every draft ships with timestamped citations and verbatim pull-quotes.

  • Built transcript ingestion via Supadata with [M:SS] markers post-processed into clickable youtube.com deep-links inside every draft
  • Engineered a voice-cloning pipeline: 1–3 user writing samples + a custom system prompt instruct Gemini to match sentence rhythm, vocabulary, and signature phrases
  • Shipped a Stripe credit-pack billing flow with HMAC-verified webhooks, atomic credit grants, Google OAuth sign-in, and a live /billing/diagnostic/ config page; deployed on Vercel with custom canonical-host middleware so OAuth survives deploy-hash URLs
DjangoPostgreSQLGemini 2.5 FlashStripeTailwind CSSVercel
Live GitHub
Project · 2025

Restaurant Web Work

Venetian Italian Eatery + Valentino's

Two restaurant engagements: a full mobile-responsive redesign of Venetian Italian Eatery's site, and frontend + Toast POS work across multiple Valentino's locations.

  • Redesigned Venetian's site for mobile + cross-device — improved page load + drove direct-order conversions
  • Maintained dynamic content (menus, promos, media) across multiple Valentino's restaurant sites
  • Configured + troubleshot Toast POS — streamlined order processing during peak hours
HTML / CSSJavaScriptToast POSResponsive design

Toolbox

Skills & stack

Languages

PythonTypeScriptJavaScriptJavaSQL

Frontend

ReactNext.jsTailwind CSSThree.jsReact Three Fiber

Backend

FastAPIDjangoNode.jsREST APIsWebSocketsStripe

Cloud & Data

Google Cloud (Cloud Run, Vertex AI, Cloud SQL, Cloud Build)RedisPostgreSQLpgvectorMySQLNeonVercelGit

AI / ML

Claude APIGemini APIGoogle ADKMulti-Agent SystemsRAGEmbeddingsKnowledge GraphsPandasNumPyScikit-learnMatplotlib

Contact

Let's build something.

Currently online

I'm open to internships for Summer 2027 and always up for collaborating on AI-grounded products and ambitious interactive web work.

  • Response

    < 24 hrs

  • Best for

    AI agents · RAG · Full-stack

Send a message

Goes straight to my inbox. I read everything.

0 / 4000
milam5@morgan.eduLinkedInGitHubBaltimore, Maryland

© 2026 Mingma Lama. All rights reserved.

Built with Next.js, React Three Fiber & Tailwind CSS.