Rishabh Garg

Primary · Footprint 5 · 7 edges

Problem

Speech emotion recognition forces a choice: a small spectrogram CNN is cheap and shallow, a large audio language model is capable and expensive. Picking one means accepting the other’s failure mode.

Approach

Build both and let them be different stages rather than competing answers. A four-layer CNN reads 128-band Mel spectrograms for the cheap pass; Qwen2-Audio-7B-Instruct, fine-tuned with LoRA, handles what the spectrogram alone cannot separate.

Outcome

The CNN reached ~82.9% validation accuracy on RAVDESS, and the fine-tuned Qwen2-Audio was evaluated on speaker-independent splits — so the number is about emotion, not about having memorised the speakers.

Flow

Data → Processing → Model → Training → Evaluation → Output

Stack

Python · TensorFlow / Keras · PyTorch · Neural Networks · Speech Processing · Transformers · Machine Learning

Primary · Footprint 5 · 7 edges

Problem

Telemedicine platforms tend to be either a video call with a booking form bolted on, or a symptom checker that confidently tells people they are fine. The first does no triage; the second does dangerous triage.

Approach

One platform carrying both, with the safety built into the model’s output rather than the copy around it: symptom-based triage and image-based skin severity classification, each returning a confidence score, each able to escalate to an emergency referral instead of an answer.

Outcome

A complete doctor–patient workflow — authentication, scheduling, real-time consultation and digital prescriptions — on a medical-record system that keeps assessments, consultations, prescriptions and transactions as structured records rather than as chat history.

Flow

Triage → Classification → Safeguards → Consultation → Records

Stack

React · Node.js · Express · MongoDB · JavaScript · WebRTC · Machine Learning

Primary · Footprint 4 · 7 edges

Problem

A Smart System Upgrade is only "hitless" if forwarding hardware has every route before traffic resumes. Downloading them as host-table entries is the slow path, and on a box carrying a full table that download time is time the switch is not forwarding the way it claims to.

Approach

The feature migrates IPv4 and IPv6 routes out of host routes into the ALPM table during the SSU boot stages, before the control plane flushes routes down to hardware — ALPM downloads to packet-forwarding silicon significantly faster than the host table does. My job was not to build it but to establish that it actually happens: that the migration fires at the right boot stage, that both address families make it across, and that it holds when the table is scaled up.

Outcome

Migration verified across the SSU boot stages for both v4 and v6 at scale, with the packet-loss window during the high-availability transition measured rather than assumed — which is the only number that says whether an upgrade was hitless.

Flow

Boot stage → Migration → Download → Flush → Verification

Stack

Arista EOS · Hardware TCAM · L2 / L3 Protocols · Packet Routing · IxNetwork (Ixia) · Wireshark · Computer Architecture

Notable · Footprint 3 · 6 edges

Problem

The control-plane CPU has finite queue capacity, and anything that can reach it can flood it. When those queues fill, the protocols keeping the switch in the network are the first thing starved — so a box under a control-plane flood does not crash, it quietly stops being a router.

Approach

Access control lists are programmed into hardware to police the CPU queues dynamically. I tested the feature three ways rather than one: at SCALE, to find where ACL and TCAM capacity actually runs out; for PERFORMANCE, to see whether policing costs anything when nothing is attacking; and with NEGATIVE cases, because a protection feature that never refuses anything has not been shown to work.

Outcome

The feature’s utility established rather than assumed — CPU queue protection holds at hardware scale, and the negative cases confirmed it drops what it is supposed to drop instead of passing everything through.

Flow

ACLs → TCAM → Flood → Scale & performance → Negative cases

Stack

Arista EOS · Hardware TCAM · L2 / L3 Protocols · IxNetwork (Ixia) · Wireshark · Computer Architecture