Bruce Cheng — software engineer. A notebook portfolio.
Bruce Cheng
notebook · 2026
MS Software Engineering · Carnegie Mellon
open me
contents
- About1
- Where I've Worked2
- What I've BuiltHelport AI3
- Robin4
- Robin: what it isn't5
- Parking Spot Locator6
- NightyNight7
- Research Assistant Agent8
- Capitawise9
- PawPrints10
- Tech Stack11
- Life & Interests12
- Let's Work Together13
scroll, drag a page corner, or use on your keyboardswipe sideways to turn the page; scroll down when there's more
about
Bruce Cheng
New-grad software engineer, available Dec 2026
MS in Software Engineering at CMU Silicon Valley. I build agentic systems, and the backends that keep them fast.
looking for
Forward Deployed EngineerAI / agent engineeringFull-stackNew-grad SWE
Bay Area · remote · open to relocate
- Helport AI — production call assistant. 6× lower inference cost, 35% shorter handling time.
- Robin — a personal AI workbench with a lasting working context, one that helps you act.
- CMU AGAI — research on detecting and explaining LLM hallucinations.
Carnegie Mellon University · Silicon ValleyMS Software Engineering · GPA 3.6 · Dec 2026
UC San DiegoBS Mathematics–Computer Science · GPA 3.6 · Mar 2025
If you do not work on an important problem, it's unlikely you'll do important work.
timeline
Where I've Worked
Sep 2026 – Present
Graduate Researcher
CMU AGAI Group
Identified and resolved flaws in LLM hallucination-detection experiments, improving detection precision by 6 percentage points through systematic evaluation and debugging.
Sep 2024 – Jun 2025
AI Product Developer / PM
Helport AI
6× lower inference cost, 35% shorter handling time. see p.3
Jul 2024 – Sep 2024
Software Engineer Intern
Convoloo
RAG medical chatbot. Response time 2.5s 1.5s.
Sep 2023 – Apr 2024
Instructional Assistant
UC San Diego · MATH 20D
Office hours, review sessions, and grading.
The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point.
6×lower inference cost — $0.002 → $0.00031 a match
- promoted to full-time, Mar 2025
- 50k+ call records a week, into dashboards
- release cycle 14d 5d
work · Sep 2024 – Jun 2025
Helport AI
AI Product Developer / PM · Helport AI
- Built Python/FastAPI services for an AI call assistant across mortgage, healthcare, insurance, and government workflows; shipped Docker/CI/CD deployments and real-time guidance APIs.
- Redesigned the intent-matching pipeline from Dialogflow to a Transformer-based system on Gemini 2.0 + Vertex AI, cutting inference cost 6× ($0.002 $0.00031 per match).
- Reduced average handling time 35% through systematic A/B testing of conversational flows and escalation triggers.
35% AHT 6× cost reduction14d 5d release cycle15% conversion
I wanted an agent that cuts the time I spend finding and filing things, so more of my day goes to what I want to do, or should be doing.
project · Aug 2026 – Present
Robin
Robin — Personal Agent Workspace · Open source · fork of agegr/pi-web
A personal AI workbench with a lasting working context — one that helps you act, not just answer.
- The problem isn't a missing chat box: information is scattered, AI conversations are cut off from the work, and picking anything back up means explaining it all again.
- Robin knows what you're working on — a chapter, a practice problem, a research topic, an opening — and helps move it forward: it explains, searches and filters, and through limited tools records tasks, schedules and updates records.
- It keeps the working state: whether a todo is done, when an interview starts, when a problem is due for review, what a research thread has collected — none of it only in a chat log.
one shared work statefour workspacesacts within set permissions
The Analytical Engine has no pretensions whatever to originate any thing. It can do whatever we know how to order it to perform.
project · Robin, continued
What it isn't
- Not just a dashboard. It doesn't stop at showing information — it helps handle it.
- Not just a chat assistant. The work state is kept apart from the conversation, so starting a new chat doesn't wipe the work.
- Not a fully autonomous agent. The AI's permissions are limited to each workspace; important decisions, and anything that needs confirming, stay with the person.
- Not a replacement for every other tool. It connects mail, courses, job sources and knowledge bases rather than rebuilding them.
- Built on Pi Web. Pi Web provides the agent sessions and the tool runtime; Robin organises the actual work on top of them.
Simplicity is prerequisite for reliability.
3×faster than traditional visual search — 45s → 15s
- sponsored by Bosch
- a plain-English query comes back as a highlighted map overlay
project · Aug 2025 – Dec 2025
Parking Spot Locator
CMU × BOSCH
- Bosch-sponsored parking locator using vision-language models and CLIP embeddings for spatial reasoning over sensor data.
- Achieved 3× speed improvement (45s 15s) over traditional visual search baselines.
- Deployed inference endpoint on AWS EC2 with FastAPI and async request queuing.
3× faster (45s 15s)
- 4 audiences: toddler · child · teen · adult
- 7 narrator voices
project · Apr 2025 – Present
NightyNight
Independent
- Full-stack AI bedtime science story generator with React/TypeScript frontend and FastAPI backend.
- LangGraph multi-agent pipeline handles story ideation, narrative generation, and age-adaptive rewriting in parallel.
- SSE streaming architecture delivers real-time story chunks to the client without polling.
LangGraph multi-agentElevenLabs TTS4 audience profiles7 narrator voices
The first principle is that you must not fool yourself — and you are the easiest person to fool.
30+academic papers indexed in Milvus
- cross-language retrieval + translation
project · Jan 2026 – Mar 2026
Research Assistant Agent
CMU 14-825
- Agentic research assistant using LangGraph for multi-step literature review, gap analysis, and summarization.
- RAG pipeline over 30+ indexed academic papers using Milvus vector database for semantic retrieval.
- Deployed on GKE with Horizontal Pod Autoscaler for concurrent multi-user workloads.
30+ papers indexedGKE auto-scaleMultilingual
Plan to throw one away; you will, anyhow.
of 20+ teams at the Franklin Templeton Hack-a-Thon · $7,000
- first-contact resolution 68% 85%
project · Mar – Jun 2024
Capitawise
Franklin Templeton Hack-a-Thon
- Led full-stack development of an AI banking chatbot for the Franklin Templeton Hackathon, placing 2nd of 20+ competing teams.
- GPT-4o dialogue layer over Node.js and Flask services behind a React client.
- Cut average response time by 40% and raised first-contact resolution from 68% to 85%.
2nd Place · $7,00040% faster replies68% 85% resolution
Talk is cheap. Show me the code.
of 30+ teams at the Franklin Templeton Hack-a-Thon · $15,000
- team of 4 — product owner and primary full-stack developer
project · Mar – Jun 2023
PawPrints
Franklin Templeton Hack-a-Thon
- Web3 pet adoption platform for Franklin Templeton Hack-a-Thon, placing 1st of 30+ teams.
- Blockchain-based pet identity and adoption history stored on-chain for immutable provenance.
- Full-stack application: React.js frontend, Express.js API, MySQL relational database.
1st Place · $15,00030+ teams
It can scarcely be denied that the supreme goal of all theory is to make the irreducible basic elements as simple and as few as possible without having to surrender the adequate representation of a single datum of experience.
- only tools that show up in the work
technologies
Tech Stack
- Languages
- PythonTypeScriptJavaScript
- Backend
- FastAPIFlaskNode.jsExpress.jsNestJSMySQL
- Cloud
- AWSGCPDockerGKETerraform
- AI
- LangGraphLangChainRAGVertex AIOpenAI APIMilvus
- Frontend
- ReactNext.jsStreamlit
ENFJ-A · 16Personalities, Jan 2023
beyond code
Life & Interests
Characters and sci-fi scenes I draw between builds.
films
games
We can only see a short distance ahead, but we can see plenty there that needs to be done.
- full-time software roles, from Dec 2026
- at CMU Silicon Valley until then
get in touch
Let's Work Together
Open to full-time roles, collaborations, and interesting conversations.
the deck
0 / 13 collected
every page has a card — read them all