Window the sequence
Overlapping windows preserve coverage across long DNA inputs.
HelixGuard
DEEP LEARNING / GENOMICSResearch capstone · private source
Read the sequence. Question the prediction.
THE CONTEXT
Assess long DNA sequences without hiding gaps in coverage.
THE KEY DECISION
Window the input; keep predictions and evidence checks separate.
THE RESULT
A research interface that reports coverage, output and unknown evidence explicitly.
HelixGuardSEQUENCE → ASSESSMENT
MECHANISM STUDY / 01SCHEMATIC
TAKE A CLOSER LOOKSELECT A STAGE ↓
SCROLL TO TRACE. SELECT TO EXPLORE.
Follow the connections as you scroll, or select a stage to see what happens there.
Learn & aggregate
A Transformer produces representations; window outputs converge into a sequence assessment.
Check the evidence
Reference-linked evidence passes through explicit biological and automata-inspired checks.
A genomic research system that keeps learned prediction, symbolic checks and supporting evidence distinct. A label is the beginning of the explanation.
PROJECT TYPE
Capstone research & engineering
FOCUS
AI / ML + Research + Backend
STATUS
Research capstone · private source
/01
The question.
A long DNA sequence rarely fits in a model’s input window. Truncating it hides what was never evaluated; a confident prediction without supporting evidence can hide even more.
/02
The approach.
Overlapping windows cover the submitted sequence. A custom k-mer Transformer produces window outputs that are aggregated into a sequence-level assessment. Reserved references supply query-linked evidence when a match exists; explicit constraints and a pushdown-automaton guardrail check that evidence separately.
/03
What came out of it.
The research interface reports coverage, model output and evidence status independently. An absent evidence match remains unknown. The portfolio illustrates this mechanism without publishing private code, data or model artifacts.
01
Coverage is part of the answer
Windowing and aggregation make the evaluated portion of a sequence explicit instead of silently truncating the input.
02
Unknown is an honest state
Evidence retrieval and learned classification answer different questions. Missing reference evidence is never presented as proof of safety.
03
A separate privacy experiment
CKKS-related linear evaluation is an experimental branch. The main model prediction path remains plaintext; the diagram keeps those boundaries visible.
Under the surface.
- Sequence modeling
- Python · PyTorch · k-mer Transformer
- Evidence & constraints
- Reference matching · biological rules · pushdown automaton
- Research interface
- FastAPI · sequence/FASTA input · structured reports
SCOPE & LIMITS
- Research software; no clinical or regulatory validation is claimed.
- The source repository and model artifacts are private.
- Experimental encrypted scoring is separate from full model inference.
NEXT PROJECT / 02
Have an idea of your own? Let’s talk