Skip to content

L0 input layer

  • x1 Retrieval
  • x2 Computer vision
  • x3 Reinforcement learning
  • x4 Production automation

Karl Neil Sequeira

I spot the gap in how work gets done, then build what bridges it.

From retrieval systems and computer vision to reinforcement learning and production automation, I build the model, the evaluation, and the product around it.

Résumé
  • AI Ops Engineer @ AHA Smart Homes
  • Chancellor’s Gold Medalist 2026
  • B.Sc. (Hons) AI & ML
  • IIT Madras T4E 2025 presenter

Don't take my word for it.

These aren’t portfolio numbers. They’re things I actually measured.

  1. 01 / 04Ask My Papers

    100%

    hit@6, 16 labelled questions

    The right passage, every time I tested it

    For all 16 questions I wrote answers for, the passage holding the answer was in the top 6 results the system retrieved.

    See how it was measured

    Ask My Papers answering a question with its cited source passage
    Real screenshot

  2. 02 / 04Ask My Papers

    0.677 → 0.783

    MRR, before and after reranking

    The right passage moved closer to the top

    Adding a second model that re-scores the results pushed the correct passage higher in the list, so the answer is built from better context.

    See how it was measured

    Chart from the measured values

  3. 03 / 04Reward Hacking in LunarLander

    0 → 29 / 50

    successful landings out of 50

    From hovering forever to landing the rocket

    My first reward taught the agent to hover and never land: 0 of 50. After I fixed the reward it landed 29 of 50 test runs.

    See how it was measured

    Real recording

  4. 04 / 04Floor Plan Intelligence

    4,035

    labelled floor-plan images

    Floor plans labelled to teach a model to read them

    The dataset I built at AHA Smart Homes, sorted into architectural, electrical, looping and unusable plans. The model is still in progress.

    See how it was measured

    Recreated HA editor with a sample floor plan, dummy data
    Recreated, dummy data

BP-01Retrieval: Ask My Papers

“I thought vector search would be enough.”

Step through what I saw, what I changed, what broke, and what fixed it.
OBSERVING

I asked a question whose answer sits in one specific passage. Vector search ranked that passage #250: all four papers are about floor plans, so their chunks sit almost on top of each other.

Key passage, vector search
rank #250
Measured, README

16 labelled questions is a small evaluation set, not a guarantee.

Step 1, Vector only: I asked a question whose answer sits in one specific passage. Vector search ranked that passage #250: all four papers are about floor plans, so their chunks sit almost on top of each other.

Diagram, not a recording. Values beside it are measured.
  1. 1. Vector only. I asked a question whose answer sits in one specific passage. Vector search ranked that passage #250: all four papers are about floor plans, so their chunks sit almost on top of each other. Key passage, vector search: rank #250 (Measured, README).
  2. 2. Add BM25. Keyword search ranked the same passage #1. The words that tell these papers apart are proper nouns, and embeddings compress exactly those away. Same passage, BM25: rank #1 (Measured, README).
  3. 3. Fusion fault. Merging both lists with Reciprocal Rank Fusion buried it again. Fusion rewards agreement, so a chunk ranked #1 by one method and #250 by the other lost to chunks both rated mediocre. Key passage after fusion: buried (Observed bug, README).
  4. 4. Fix. Each method gets reserved slots before fusion fills the rest. A cross-encoder reranks 20 candidates down to 6, and a relevance gate refuses when nothing clears the bar. hit@6: 100% (Measured, 16 labelled questions). MRR: 0.677 → 0.783 (Measured, after reranking). Out-of-scope question: refused in ~1 s, not ~6 s (Measured, README).

The full write-upThe code

BP-02Reward: LunarLander

“I gave the model the wrong incentive.”

The retrieval bug was the data hiding the answer. This one was my own objective rewarding the wrong thing.
OBSERVING

Plain PPO trained for 1M steps, as the reference.

Mean score
142.5
Measured, 50 landings
Solved (200+)
18 / 50
Measured, 50 landings

Every version is scored on the game’s own reward over the same 50 landings. One training run per version, so this is evidence, not proof.

Step 1, Baseline: Plain PPO trained for 1M steps, as the reference.

Real episodes: random agent, baseline, hacked reward (hovers), fixed reward.
Real recording. The values beside it are measured.
  1. 1. Baseline. Plain PPO trained for 1M steps, as the reference. Mean score: 142.5 (Measured, 50 landings). Solved (200+): 18 / 50 (Measured, 50 landings).
  2. 2. Time fee. A fee of 0.1 for every frame pushes it to land sooner. This version solved the most landings. Mean score: 195.8 (Measured, 50 landings). Solved (200+): 37 / 50 (Measured, 50 landings).
  3. 3. Reward hack. Wait. Why is it hovering? A bonus of +1 for every upright, slow frame is a reward it can farm, so it hovered until the 1,000-frame limit. It optimised the reward, not the task. Mean score: −41.7 (Measured, 50 landings). Solved (200+): 0 / 50 (Measured, 50 landings). Timed out hovering: 32 / 50 (Measured, 50 landings).
  4. 4. The fix. I fixed the objective: potential-based shaping pays for getting closer to landing, not for staying upright. The exploit is gone, though the plain time fee still landed more often (37 / 50). Mean score: 178.1 (Measured, 50 landings). Solved (200+): 29 / 50 (Measured, 50 landings).

The full write-upThe code

02About

Behind the work.

I like working on the part of AI that happens after the demo.

The problems I find interesting are usually messy: documents that don’t retrieve cleanly, models that exploit the metric they’re given, floor plans that don’t look like clean datasets, and internal software that has to survive real users.

That’s why most of my projects involve more than a model. I build the data pipeline, the evaluation, the interface and the failure handling around it.

Right now I’m an AI Ops Engineer at AHA Smart Homes.

Off the clock

When I’m not working, I’m usually on a football pitch or a badminton court (pickleball, lately), out on a long bike ride with friends, or DJing a bit of house, tech house and pop.

AHA Smart Homes
AI Ops Engineer
Atlas SkillTech University
B.Sc. (Hons) AI & ML, 2022–2026
Chancellor's Gold Medal
2026
IIT Madras T4E
Research presenter, 2025
GPA
8.29 / 10
Awards, recognition and background

Where I've worked

  • AHA Smart Homes AI Ops EngineerMar 2026 – Present
  • Unreal Dimension Generative AI Intern (part-time)Feb 2025 – Aug 2025
  • EdTech Society Machine Learning InternJan 2025 – Apr 2025
  • Digirise Infolabs Machine Learning InternOct 2023 – Nov 2023

03Built and broken

Things I've built. And things I've broken.

BP-03Proposal: Floor Plan to Proposal

“I took proposals out of Excel.”

OBSERVING

Consultants assembled the customer proposal document by hand in Excel, about 1.5 hours each (an estimate), copying from old ones.

Proposal document
Built by hand
Before the PMT

The time is an estimate, not a measurement.

Step 1, Before: Consultants assembled the customer proposal document by hand in Excel, about 1.5 hours each (an estimate), copying from old ones.

Recreated with dummy data. Diagram, not a recording.
  1. 1. Before. Consultants assembled the customer proposal document by hand in Excel, about 1.5 hours each (an estimate), copying from old ones. Proposal document: Built by hand (Before the PMT).
  2. 2. Generate. Now the consultant adds devices and groups in the PMT, and the branded, room-by-room document is generated from that data. Proposal document: About 15 min (Estimate).
  3. 3. The gap. Project managers still re-placed every device on the electrical plan by hand. Consultants and PMs work from different offices, so devices sometimes landed in the wrong spot, like a keypad on the wrong side of the sofa. Devices placed: Twice (By the consultant, then the PM).
  4. 4. Bridge. Consultants place the devices themselves in the HA Editor, room by room, so what reaches project managers after onboarding is already right. Devices placed: Once (By the consultant).
See the system

04Every project

All 17 projects, one layer.

Filter by area. Each bar shows whether it shipped or is still in progress.

17 projects shown

n01

Ask My Papers

A learning project: a local RAG over floor-plan research papers that cites its sources and refuses to answer when the evidence is weak.

ShippedProject

n04

Facial Emotion Recognition

A passion project in computer vision: a CNN that sorts face images into seven emotions, built in a team of three.

ShippedProject

n07spotted the gap

Payment Reminders

One of my early projects at AHA, and how I learned the Periskope (WhatsApp) API.

ShippedWhere I started

n10

EP Image Editor

Electrical plans often arrived as PDFs, and they did not go through properly: consultants could not pick the page or the section of the plan they needed.

ShippedFloor plan to proposal

n13spotted the gap

WhatsApp Support Ticket Sync

WhatsApp support tickets from Periskope flow into the project tool automatically, skipping any already imported.

ShippedSupport operations

n16

WhatsApp Project Groups

Consultants or someone from operations created each project’s WhatsApp group by hand, and it was often unclear who to add: which project manager, which coordinator.

ShippedIntegrations and internal tools

n02

Reward Hacking in LunarLander

A learning project after playing LunarLander: I trained an RL agent to land the rocket, rewarded it for staying afloat, and it learned to hover instead of land. That is how I met reward hacking.

ShippedProject

n05

Pneumonia Detection from Chest X-rays

A college project: a CNN that flags pneumonia in chest X-rays, built in a team of three on a public Kaggle dataset.

ShippedProject

n08spotted the gap

Floor Plan to Proposal

Replaced the customer proposal document consultants assembled by hand in Excel with one the PMT generates: a branded, room-by-room PDF (about 1.5 hours down to about 15 minutes, estimated). Then gave them an editor to place every device on the plan themselves.

ShippedFloor plan to proposal

n11spotted the gap

Additional Items Approval Workflow

Add-ons requested on site go to the customer for email approval, then append to the sales order with a record of who approved what.

ShippedApprovals and sales orders

n14

Support Dashboard and Reports

The support team wanted to see how many tickets were raised, how many were solved and how long each took, instead of reading tickets one by one.

ShippedSupport operations

n17

Website Deployment and GIS Uploads

A fellow developer who runs the company website and a BI tool found uploading to S3 and pushing changes live slow and fiddly.

ShippedIntegrations and internal tools

n03

Predictive Modelling Notebooks

EDA and baseline models on four public datasets, from my first ML internship. Outputs shown exactly as the notebooks printed them.

ShippedProject

n06

Zoho CRM Lead Dashboard

My first project at AHA, requested by the marketing head, who needed to see where leads were coming from and which ads brought them in.

ShippedWhere I started

n09spotted the gap

Floor Plan Intelligence

The goal: upload any floor plan and have a model find the walls, doors and windows, then place the proposal’s devices automatically. Step one is a labelled dataset; the model is still being built.

In progressFloor plan to proposal

n12spotted the gap

GST Verification

Company GST details on forms were typed by hand and never checked.

ShippedApprovals and sales orders

n15

Zoho Inventory Sync

Inventory data lived only in Zoho, out of reach of internal dashboards, and Zoho’s standard plan allows 2,500 API calls a day.

In progressIntegrations and internal tools

Open the full project index

05At work

AHA Smart Homes: AI and software systems used inside a real business.

AI Ops Engineer, Mar 2026 – Present. AHA designs and installs home automation. I build features for its internal Project Management Tool (PMT) used by consultants, project managers and the support team, and I am working on computer vision for reading electrical floor plans.
  1. 01Lead comes inSolved by creating a lead dashboard
  2. 02Consultant meets the client
  3. 03Proposal documentSolved by generating the document from the PMT
  4. 04Devices on the planSolved by building the HA Editor
  5. 05Project groupSolved by auto-creating the WhatsApp group
  6. 06Onboarding and paymentSolved by automating reminders and checking GST
  7. 07Sales order and add-onsSolved by adding customer approval for add-ons
  8. 08Installation on site
  9. 09SupportSolved by syncing tickets and adding a dashboard

Lit stages are where I built something. Open one to see it.

AHA Smart Homes screens on this site are recreated with dummy data.

06The toolbox

What I use to build the systems above.

karl.skills.embed() → move to querytop-5:

Illustrative layout grouped by skill area, not a trained embedding.

  • I build models: Python, PyTorch, TensorFlow, Keras, OpenCV, scikit-learn, NumPy, pandas
  • I build retrieval systems: BM25, Vector search, RRF, Cross-encoder reranking, Embeddings, Chroma, LangChain, Ollama
  • I build products around them: React, Vite, Tailwind, Next.js, FastAPI, MongoDB, Postgres, Supabase, Docker, AWS, GCP, Azure
  • I evaluate them: pytest, Retrieval evaluation, Experiment design, Failure analysis
  • Python
  • PyTorch
  • TensorFlow
  • Keras
  • scikit-learn
  • OpenCV
  • NumPy
  • pandas
  • RAG
  • BM25
  • Vector search
  • RRF
  • Cross-encoder reranking
  • LangChain
  • Ollama
  • Chroma
  • Embeddings
  • React
  • Vite
  • Tailwind
  • Next.js
  • FastAPI
  • Supabase
  • Postgres
  • MongoDB
  • AWS
  • GCP
  • Azure
  • Docker
  • pytest
  • Experiment design
  • Retrieval evaluation
  • Failure analysis
  • Data pipelines
  • API integration
  • Automation

I build models

Image models and RL agents, from preprocessing to evaluation.

See it in Facial Emotion Recognition

I build retrieval systems

Hybrid search with a reranker and a refusal gate, measured on labelled questions.

See it in Ask My Papers

I build products around them

The editors, dashboards, syncs and APIs people use around the model.

See it in Floor Plan to Proposal

I evaluate them

Labelled eval sets, regression tests written against real bugs, and failure analysis.

See it in Reward Hacking in LunarLander

07Field notes

Field notes.

  1. Retrieval, 1 min readWhy my RAG pipeline needed BM25Four papers on the same topic sat on top of each other in embedding space. Keyword search found what vectors ranked #250.
  2. Reinforcement learning, 1 min readWhat reward hacking taught me about RLA bonus for staying upright and slow taught a PPO lander to hover until time ran out: 0 landings out of 50.
  3. Systems, 1 min readHow I designed ownership rules for a WhatsApp → PMT syncSyncing support tickets between two systems turned out to be an ownership problem, not an API problem.
  4. Computer vision, 1 min readWhy I don’t publish an accuracy number yetThe floor-plan model at AHA is still training. Here is what exists instead: a labelled set of 4,035 images.

08Say hello

Feel free to reach out.

Whether it’s an idea you’re working on, a question about something here, or just to say hi, I’m always happy to chat.
  • AI/ML engineering
  • Computer vision
  • Retrieval and LLM systems
  • Applied AI
  • AI engineering
karlsequeira2004@gmail.com

Email meLinkedInGitHubRésumé