I. RESEARCH II. DOSSIER III. NOTES
S+KHAVIN r 007*
Independent Research Beyond The Known™
Vision, Interpretability, Efficiency ♔

RESEARCH LOG

Aug 2026
K Heads is All You Need task-dependent circuit granularity and zero-training routing in LLMs.
↗ OPEN
Aug 2026
The Shape of Computation language models build universal, conserved task geometries that are causally inert.
↗ OPEN
Jul 2026
Necessary But Not Sufficient convergent computational geometry in the transformer residual stream.
↗ OPEN
Jul 2026
The HeadGenome Taxonomy structural and causal map of attention head specialization across 1,568 heads in 4 architectures.
↗ OPEN
May 2026
O(1) Decode Step Attention training-free proactive KV cache eviction for any transformer.
↗ OPEN
Mar 2026
Trinetra proactive temporal event graph for long-form video understanding.
↗ OPEN
2024
Team Lead at Google GenAI Hackathon Top Project Nationwide.
↗ OPEN
2024
Selected for AlgoUniversity Technology Fellowship top 5% of 20,000 applicants.
↗ OPEN
2023
Udacity Next Gen Tech Booster Scholarship 500 of 17,000 global applicants.
↗ OPEN

About Me

Indian Space Force Top Secret

Classified Dossier - S Khavin 7

+About

© About エクスペリエンス (Top Secret -003) Built with Coffee(Lots of it!!)

✯ NAME: S. KHAVIN

✯ RESEARCH: MECHANISTIC INTERPRETABILITY • EFFICIENT INFERENCE

✯ EDUCATION: B.E. CSE, SATHYABAMA • 2023 - 2027

✯ STATUS: ACTIVE


✯ RESEARCH SUMMARY:
INDEPENDENT RESEARCHER IN TRANSFORMER INTERPRETABILITY AND INFERENCE EFFICIENCY.
FOCUS: CAUSAL, NEUROSCIENCE-INSPIRED METHODS TO UNDERSTAND & ACCELERATE MODELS.
SIX FIRST-AUTHOR PREPRINTS SPANNING GEOMETRY, ROUTING, KV-CACHE, AND VIDEO UNDERSTANDING.


✯ PREPRINTS:

• K HEADS IS ALL YOU NEED: CIRCUIT GRANULARITY & ZERO-TRAINING ROUTING.
• THE SHAPE OF COMPUTATION: GEOMETRIC TRAJECTORIES IN RESIDUAL STREAMS.
• NECESSARY BUT NOT SUFFICIENT: CAUSAL TESTS OF UNIVERSAL GEOMETRY.
• HEADGENOME: CAUSAL MAP OF 1,568 ATTENTION HEADS ACROSS 4 ARCHITECTURES.
• O(1) DECODE STEP ATTENTION: PROACTIVE KV-CACHE EVICTION.
• TRINETRA: O(N) VIDEO INGEST, O(1) QUERY TEMPORAL EVENT GRAPH.

✯ DISTINCTIONS:

•  SWE CO-OP INTERN @ ALGOUNIVERSITY | UDACITY SCHOLAR (1 OF 500)
•  ALGOUNIVERSITY TECHNOLOGY FELLOWSHIP, TOP 5% OF 20,000

✯ METHODS:
PYTORCH • FAISS • CAUSAL PATCHING • DIRECT LOGIT ATTRIBUTION • CKA • DTW

ACT III // ARCHIVAL RECORD

Research Archive

Areas: Mechanistic Interpretability, Vision, Neuroscience

PREPRINTS & CIRCUITS (06)
EXPERIENCE & POSTS (10)
  • AI & Automation Intern Chennai Airport
  • AI & FullStack Intern AlgoUniversity (YC)
  • Business Analyst Intern (Hero) Business.io IITM
  • Artificial Intelligence Intern AICTE
  • Sponsorship & Collab Head LingoLabs IITM
  • ML Research Intern (Co-Lead) Omdena
  • Software Engineer Intern Headstarter (YC)
  • S5 Nights & Weekends Fellow Buildspace (YC)
  • Technical Team Member GDSC SIST
  • Web Development Specialist Philosophy Club
AWARDS & HONORS (04)
  • Udacity Next Gen Tech Booster Scholarship Top 500 of 17,000 Global Applicants
  • AlgoUniversity Technology Fellowship (ATF) Top 5% of 20,000 Applicants Nationwide
  • National Entrepreneurship Challenge (E-Cell) 5th Place Nationwide // IIT Bombay
  • Technical Architecture for University Clubs Pod SIST · Alpha Gamma Nu IITM · GDSC
GET IN TOUCH
skhavinshankar@gmail.com CHENNAI, TAMIL NADU, IN (IST)
S. Khavin
S. KHAVIN DOSSIER 007
007 // CINEMATOGRAPHIC REEL ARCHIVE

Projects

CLASSIFIED EMPIRICAL SYSTEMS // 35MM OPTICAL FILM REEL

KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 01
K Heads is All You Need
EXP 01
24 FPSEMULSION 007▸ 01A
KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 02
The Shape of Computation
EXP 02
24 FPSEMULSION 007▸ 02A
KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 03
Necessary but Not Sufficient
EXP 03
24 FPSEMULSION 007▸ 03A
KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 04
The HeadGenome Taxonomy
EXP 04
24 FPSEMULSION 007▸ 04A
KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 05
O(1) Decode Step Attention
EXP 05
24 FPSEMULSION 007▸ 05A
KODAK 5219SAFETY FILMEASTMAN 5219 ▸ 06
Trinetra: Temporal Event Graph
EXP 06
24 FPSEMULSION 007▸ 06A
FRAME 01 // 06•CAUSAL CIRCUITS // ROUTING•TOP SECRET // LVL 6

K Heads is All You Need

Task-Dependent Circuit Granularity & Zero-Training Routing in LLMs

Needle-in-a-haystack retrieval supported by a 6-head causal circuit; zero-training geometric router on NVIDIA RULER 4K.

ACT IV // THEORETICAL INQUIRIES & ESSAYS

Field Dispatches

NOTE 01 // ESSAY THEORETICAL FRAMEWORK

The Trunk and Branch Theory of Computation

Large language models do not compute through isolated lookup tables; they construct universal, conserved task manifolds in activation space. Categories diverge progressively like tree branches from a shared semantic trunk, passing temporal-shuffle and category-confusion null controls.

READ ESSAY →
CENTRAL QUESTIONS ACTIVE INVESTIGATION
01
Universal Representation Geometry

Do diverse architectures build conserved task manifolds that are causally load-bearing, or are geometries epiphenomenal byproducts?

02
Causal Circuit Sparsity Boundary

Why is retrieval compressible to K=6 attention heads while algorithmic reasoning strictly resists sparsification?

03
Proactive Inference Memory

Eliminating quadratic O(N) context scaling by predicting and bounding attention allocations to O(1) before query arrival.

06 PREPRINTS
/
1,568 HEADS MAPPED
/
O(1) DECODE COMPLEXITY
/
007 AGENT CLEARANCE
©2026 S. KHAVIN | ALL RIGHTS RESERVED
CHENNAI, INDIA (IST) / CLEARANCE: LEVEL 6