Data Analyst Intern
Bluestock Fintech
Built a mutual-fund analytics platform covering 40 schemes, 10 datasets, 87K+ records, and 4.5 years of NAV history.
Engineer / builder / research explorer
Apoorv Pal
Software engineer
New Delhi, India
I build intelligent systems that survive beyond the demo.
B.Tech Software Engineering student at DTU. I work where AI behavior meets data reliability, product constraints, and the awkward edge cases that make software real.
- 01Currently exploring
- Reliable retrieval, evaluation, and data infrastructure
- 02Latest build
- ConceptGraph AI
- 03Open to
- Engineering internships and research collaborations
- 04Last updated
- July 2026
About
Person behind the systems
Building is how I learn what the system was hiding.
Apoorv is pursuing a B.Tech in Software Engineering at Delhi Technological University, with an 8.655/10 CGPA. His work spans AI systems, data infrastructure, full-stack products, and applied machine learning.
He enjoys moving from a rough idea to a working implementation, then returning to the states, constraints, and failure paths the first version ignored. He has solved 500+ algorithmic problems and placed seventh at HackOrbit 2025.
C++PythonTypeScriptReactFastAPIPostgreSQLNeo4jQdrantDockerPandasScikit-learnSupabase
Where the work met constraints.
Open Source Contributor
Enginow / ELUSOC
Contributed through the Enginow Open Source Program and earned Iron Developer recognition with 315 contribution points in ELUSOC 2026.
Learning in commits, constraints, and shipped work.
- 01
Software engineering at DTU
Started turning coursework into shipped systems and disciplined problem-solving practice.
- 02
HackOrbit · seventh place
Moved from solo building into time-boxed product and engineering collaboration.
- 03
ConceptGraph AI
Built a citation-grounded GraphRAG system spanning ingestion, vector search, graph expansion, and reranking.
- 04
Bluestock + open source
Worked with 87K+ financial records and contributed through Enginow / ELUSOC.
Selected systems / 01—04
Projects built past the happy path.
Compact records here. Architecture, failures, and engineering decisions live inside each project page.
Knowledge system / GraphRAG
ConceptGraph AI
An academic GraphRAG platform that turns course PDFs into a queryable concept graph and returns reranked answers with page-level citations.
500+Concept nodes rendered
Privacy-aware wellbeing product
Serenity AI
A mood-aware journaling and conversation product designed around secure personal data, recent context, and restrained AI assistance.
6Message context bound
Machine learning / race simulation
F1 Race Strategy Decision Support
A decision-support application that combines lap-time prediction with Monte Carlo simulation to compare Formula 1 pit-stop strategies across real 2021–2025 race data.
1.41sHGB MAE
Constraint system / data engineering
Smart Classroom & Timetable Scheduler
A scheduling system that models rooms, faculty, courses, time windows, and conflicts as explicit constraints rather than spreadsheet exceptions.
Workbench
Smaller investigations that sharpen the larger systems.
W.01
Resume ATS Analyzer
Local browser parsing, five-dimension ATS scoring, job-description matching, and downloadable feedback.
React / TypeScriptW.02
Mutual Fund Analytics
40 schemes, 87K+ records, and 4.5 years of NAV history.
Bluestock / 2026W.03
Phishing Website Detection
EDA, preprocessing, baseline models, Optuna tuning, and saved inference artifacts on the PhiUSIIL dataset.
Python / MLField notes / Issue 03Things learned
Things learned
the expensive way.
Short essays on retrieval, product behavior, and systems that have to operate under partial failure.
Why vector retrieval alone was insufficient
Semantic similarity finds nearby language. It does not automatically preserve prerequisites, hierarchy, or evidence chains.
Read field note
A useful learning system has to recover more than a similar paragraph. It needs to understand where a concept sits, what it depends on, and which source supports the answer.
That changed the retrieval design: vector search became a candidate generator, while graph expansion, reranking, and citation checks decided what could reach the final context.
Designing document-processing state machines
A pipeline is easier to trust when every transition has an owner, an invariant, and a cleanup path.
Read field note
UPLOAD, VALIDATING, PROCESSING, READY, and FAILED are product states, not merely backend details. Retrieval should never see half-written data.
Retries must be idempotent, duplicate inputs need stable identities, and a failed stage should remove or quarantine its partial writes before the next attempt.
Why successful demos still fail in production
The happy path demonstrates possibility. Production is the accumulation of unusual inputs, interrupted requests, and ambiguous ownership.
Read field note
The real system begins where the demo ends: timeouts, rate limits, malformed documents, stale sessions, and retries after partial success.
I now treat failure behavior as part of the feature specification. If the system cannot explain its state or recover safely, the workflow is not finished.
Optimistic UI beyond perceived speed
Optimism is a temporary claim the interface makes. Good product engineering defines how that claim can be corrected.
Read field note
A fast interface should still make pending, confirmed, and rejected states legible. Rollback cannot feel like data silently disappearing.
Optimistic flows work best when mutations have stable client IDs, clear reconciliation rules, and copy that communicates uncertainty without creating anxiety.
Open channel / New Delhi
Have a difficult system worth building?
I’m interested in AI infrastructure, full-stack product engineering, data-heavy systems, and research work where reliability matters.
Copy it for later, or open your mail client directly. Résumé available as web and PDF editions.