Skip to main content
JOTARA LAB AI & Security Software Studio
Systems Verification Sandbox

Active Lab Prototypes & Technical Case Studies.

Active R&D projects from Jotara Lab, plus technical case studies of production-grade systems we have built. Architecture details and stack choices are public.

The Challenge: Sending proprietary chemical formulas and clinical trial records to public LLM APIs exposed Adara Pharma to intellectual property leaks and regulatory risk.The Solution: We deployed quantized Gemma and Llama models inside Adara's isolated VPC, connected to a retrieval-augmented generation (RAG) pipeline over their internal research repository.

METRIC // 01
14+ Hours/Employee/WeekStaff Time Saved
METRIC // 02
0% (Completely Sandboxed)Data Leak Risk
METRIC // 03
Air-Gapped Private SubnetNetwork Isolation
METRIC // 04
Private VPC GemmaModel Core Node
// Architecture Deep Dive//HARDENED_AIRGAP
Vector EmbeddingsSelf-hosted Qdrant + BGE-M3 embeddings, 1024-dim vectors over a 2.4M-chunk internal knowledge base.
RAG PipelineLangChain orchestration with custom re-ranking; top-k=8 with metadata filters on document version.
Model QuantizationINT4-quantized Gemma 12B on a single NVIDIA A10G; vLLM inference server over VPC peering.
VPC IsolationPrivate subnet, no public ingress; egress routed through Squid proxy with domain allow-list only.

The Challenge: Internal microservices currently share a flat VPC trust boundary. A compromised pod can enumerate every peer.The Solution: An L7 reverse proxy (Envoy + custom Go filters) that issues short-lived mTLS certs per request via SPIFFE, denies lateral movement, and emits signed audit envelopes to Loki.

METRIC // 01
+12ms p99Auth Latency
METRIC // 02
Blocked by DefaultLateral Movement
METRIC // 03
SPIFFE/SPIRE issuedIdentity
METRIC // 04
Append-only, signedAudit Trail
// Architecture Deep Dive//ENVOY_SPIFFE_ENGINE
Workload Identity (SPIFFE/SPIRE)Short-lived X.509 SVID certificates rotated every 60 minutes via SPIRE agent node attestation.
L7 Reverse Proxy EngineEnvoy proxy daemon with custom Go filters enforcing fine-grained RBAC and per-route rate limits.
Lateral Movement DenialStrict default-deny egress policies; compromised pods cannot discover or query adjacent services.
Signed Audit TelemetryEvery intercepted request emits a cryptographically signed HMAC envelope streamed asynchronously to Grafana Loki.

The Challenge: Off-the-shelf prompt-injection classifiers (Llama Guard, PromptGuard) miss 15-20% of obfuscated payloads in our internal red-team tests.The Solution: A Rust pre-processor that runs regex, structural AST checks, and embedding-similarity scoring, then routes borderline inputs to a local LLM judge. The full pipeline adds 38ms median latency end-to-end.

METRIC // 01
97.4% (vs 81% baseline)Detection Rate
METRIC // 02
0.8% on clean promptsFalse Positives
METRIC // 03
12k tokens/secThroughput
METRIC // 04
+38ms p50Latency Add
// Architecture Deep Dive//AST_SANITIZER_FLOW
Rust Regex & AST SanitizerZero-allocation Rust pre-filter scanning token trees for system prompt leak attempts and jailbreak delimiters in < 4ms.
Semantic Embedding ScorerLocal BGE-micro vector similarity matching against a curated adversarial prompt embeddings vector index.
Local Arbiter LLM JudgeQuantized 3B classifier arbitrates ambiguous edge-case inputs without leaking payloads to public endpoints.
Latency Budget & FallbackStrict 40ms SLA timeout with fail-secure quarantine behavior and structured security telemetry alerts.
Direct Senior Engineering Access // Response < 2 hrs

Ready to Architect Resilient Systems?

Discuss your distributed architectures, private AI pipelines, or high-concurrency systems directly with our senior engineering team without procurement friction.

Start Consultation→Chat on WhatsApp